Fourier Analysis on Finite Groups with Applications in Signal Processing and System Design
Radomir S. Stankovid Claudio Moraga Jaakko T. Astola
IEEE PRESS
A JOHN WILEY & SONS, MC., PUBLICATION
This Page Intentionally Left Blank
Fourier Analysis on Finite Groups with Applications in Signal Processing and System Design
IEEE Press
445 Hoes Lane Piscataway, NJ 08854 IEEE Press Editorial Board Stamatios V. Kartalopoulos, Editor in Chief M. Akay J. B. Anderson R. J. Baker J. E. Brewer
M. E. El-Hawary R. Leonardi M. Montrose M. S. Newman
F. M. B. Periera C. Singh S . Tewksbury G. Zobrist
Kenneth Moore, Director of Book and Information Services (BIS) Catherine Faduska, Senior Acquisitions Editor Anthony VenGraitis, Project Editor
Fourier Analysis on Finite Groups with Applications in Signal Processing and System Design
Radomir S. Stankovid Claudio Moraga Jaakko T. Astola
IEEE PRESS
A JOHN WILEY & SONS, MC., PUBLICATION
Copyright 0 2005 by the Institute of Electrical and Electronics Engineers, Inc. All rights reserved. Published by John Wiley & Sons, Inc., Hoboken, New Jersey Published simultaneously in Canada.
No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvcrs, MA 01923, (978) 750-8400, fax (978) 646-8600, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 11 1 River Street, Hoboken, NJ 07030, (201) 748-601 I , fax (201) 748-6008 or online at http://www.wiley.comlgo/permission. Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales matcrials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. For general information on our other products and services or for technical support, please contact our Customer Care Department within the U.S. at (800) 762-2974, outside the U.S. at (3 17) 572-3993 or fax (317) 572-4002.
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic format. For information about Wiley products, visit our web site at www.wiley.com.
Library of’Congress Cataloging-in-Publication is available. ISBN-13 978-0-471-69463-2 ISBN-10 0-471-69463-0
Printed in the United States of America 1 0 9 8 7 6 5 4 3 2 1
Preface
We believe that the group-theoretic approach to spectral techniques and, in particular, Fourier analysis, has many advantages, for instance, the possibility for a unified treatment of various seemingly unrelated classes of signals. This approach allows to extend the powerful methods of classical Fourier analysis to signals that are defined on very different algebraic structures that reflect the properties of the modelled phenomenon. Spectral methods that are based on finite Abelian groups play a very important role in many applications in signal processing and logic design. In recent years the interest in developing methods that are based on Finite non-Abelian groups has been steadily growing, and already, there are many examples of cases where the spectral methods based only on Abelian groups do not provide the best performance. This monograph reviews research by the authors in the area of abstract harmonic analysis on finite non-Abelian groups. Many of the results discussed have already appeared in somewhat different forms in journals and conference proceedings. We have aimed for presenting the results here in a consistent and self-contained way, with a uniform notation and avoiding repetition of well-known results from abstract harmonic analysis, except when needed for derivation, discussion and appreciation of the results. However, the results are accompanied, where necessary or appropriate, with a short discussion including comments concerning their relationship to the existing results in the area. The purpose of this monograph is to provide a basis for further study in abstract harmonic analysis on finite Abelian and non-Abelian groups and its applications.
V
vi
PREFACE
V
Chapter 5
> Fitrictroriol E\pres\ror7s
or1
Qiiotrr 171onGroirps
Fig. 0.7
Relationships among the chapters.
The monograph will hopefully stimulate new research that results in new methods and techniques to process signals modelled by functions on finite non-Abelian groups. Fig. 0.1 shows relationships among the chapters. RADOMIRS. STANKOVIC, CLAUDIOMORAGA,JAAKKOT. ASTOLA NiS Dorrrniind. Tirrrpere
Acknowledgments
Prof. Mark G. Karpovsky and Prof. Lazar A. Trachtenberg have traced in a series of publications chief directions in research in Fourier analysis on finite non-Abelian groups. We are following these directions in our research in the area, in particular in extending the theory of Gibbs differentiation to non-Abelian structures. For that, we are very indebted to them both. The first author is very grateful to Prof. Paul L. Butzer, Dr. J. Edmund Gibbs, and Prof. Tsutomu Sasao for continuous support in studying and research work. The authors thank Dragan JankoviC of Faculty of Electronics, University of NiS, Serbia, for programming and performing the experiments partially reported in this monograph. A part of the work towards this monograph was done during the stay of R. S. StankoviC at the Tampere International Center for Signal Processing (TICSP). The support and facilities provided by TICSP are gratefully acknowledged.
R.S.S.,C.M, J.T.A.
Vii
This Page Intentionally Left Blank
Contents
Preface Acknowledgments A cronyms
I
V
vii xxiii
Signals and Their Mathematical Models 1.1 Systems 1.2 Signals 1.3 Mathematical Models of Signals References
2 Fourier Analysis 2. I Representations of Groups 2.1. I Complete reducibility 2.2 Fourier Transform on Finite Groups 2.3 Properties of the Fourier transform 2.4 Matrix interpretation of the Fourier transform on jinite non-Abelian groups 2.5 Fast Fourier transform on jinite non-Abelian groups References
11 12 13 18 23
26 28 35 ix
x
3
4
CONTENTS
Matrix Interpretation of the FFT 3. I Matrix interpretation of FFT on jinite non-Abelian groups 3.2 Illustrative examples 3.3 Complexity of the FFT 3.3. I Complexity of calculations of the FFT 3.3.2 Remarks on programming implementation of FFT 3.4 FFT through decision diagrams 3.4. I Decision diagrams 3.4.2 FFT on Jinite non-Abelian groups through DDs 3.4.3 MTDDs for the Fourier spectrum 3.4.4 Complexity of DDs calculation methods References Optimization of Decision Diagrams 4. I Reduction Possibilities in Decision Diagrams 4.2 Group-theoretic Interpretation of DD 4.3 Fourier Decision Diagrams 4.3. I Fourier decision trees 4.3.2 Fourier decision diagrams 4.4 Discussion of DiJjrerent Decompositions 4.4. I Algorithm for optimization of DDs 4.5 Representation of Two-Variable Function Generator 4.6 Representation of adders by Fourier DD 4.7 Representation of multipliers by Fourier DD 4.8 Complexity of FNADD 4.9 Fourier DDs with Preprocessing 4.9. I Matrix-valued functions 4.9.2 Fourier transform for matrix-valued functions 4. I0 Fourier Decision Trees with Preprocessing 4. I I Fourier Decision Diagrams with Preprocessing 4.12 Construction of FNAPDD 4.13 Algorithm for Construction of FNAPDD 4.13. I Algorithm for representation 4.14 Optimization of FNAPDD References
37 38 41 59 62 66 66 66 68 76 76 80 85 86 93 96 96 107 108 110
I I0
114 117 123 129 129 130 135 136 137 151 152 153 154
CONTENTS
Xi
5 Functional Expressions on Quaternion Groups 5.I Fourier expressions on finite dyadic groups 5. I . I Finite dyadic groups 5.2 Fourier Expressions on Q 2 5.3 Arithmetic Expressions 5.4 Arithmetic expressionsfrom Walsh expansions 5.5 Arithmetic expressions on Q z 5.5.I Arithmetic expressions and arithmetic-Haar expressions 5.5.2 Arithmetic-Haar expressions and Kronecker expressions 5.6 Diflerent Polarity Polynomial Expressions 5.6. I Fixed-polarity Fourier expansions in C(Q2) 5.6.2 Fixed-polarity arithmetic-Haar expressions 5.7 Calculation of the arithmetic-Haar coeficients 5.7. I FFT-like algorithm 5.7.2 Calculation of arithmetic-Haar coeflcients through decision diagrams References
157 158 158 158 160 161 163
6 Gibbs Derivatives on Finite Groups 6.I Definition and properties of Gibbs derivatives on finite non-Abelian groups 6.2 Gibbs anti-derivative 6.3 Partial Gibbs derivatives 6.4 Gibbs diflerential equations 6.5 Matrix interpretation of Gibbs derivatives 6.6 Fast algorithms for calculation of Gibbs derivatives on finite groups 6.6. I Complexity of Calculation of Gibbs Derivatives of Gibbs derivatives through DDs Calculation 6.7 6.7. I Calculation of partial Gibbs derivatives References
183
7 Linear Systems on Finite Non-Abelian Groups 7.I Linear shift-invariant systems on groups 7.2 Linear shift-invariant systems on Jinite non-Abelian groups
166 166 167 168 169
I 72 I 72
I 74 180
I 84 186 187 189 190
192
I 98 201 203
207 21 1 211 213
xii
CONTENTS
7.3
Gibbs derivatives and linear systems 7.3.I Discussion References
214 215 21 7
22 1 8 Hilbert Transform on Finite Groups 8.1 Some results of Fourier analysis onJinite nun-Abelian 223 groups 8.2 Hilbert transform on jnite nun-Abelian groups 227 23 I 8.3 Hilbert transform in jnite Jields 234 References Index
235
List of Figures
0. I
Relationships among the chapters.
2. I
FFT on the quaternion group Q2.
32
2.2
Flow-graph for FFT algorithm for the inverse Fourier transform on Q8.
33
2.3
FFT on the dyadic group of order 8.
34
3. I
Structure of the flow-graph of the FFT on the group G2x8.
46
3.2
Structure of the flow-graph for FFT on the group GS2.
48
3.3
Structure of the flow-graph for FFT on the group G32 through a part of fast Walsh transform.
49
3.4
Structure of the flow-graph for FFT on the group G32 using FFT on Q2.
50
3.5
Structure of the flow-graph for FFT on the group GGx6.
53
3.6
Structure of the flow-graph for FFT on the group G 3 x ~ .
57
3.7
Structure of the flow-graph for FFT on G24
58
3.8
Structure of the flow-graph for FFT on S3.
60
3.9
Structure of the flow-graph for FFT on G24 with FFT on S3.
61
vi
xiii
xiv
LIST OF FIGURES
3. I0 Number of operations in FFT.
63
3.11 Time requirements.
65
3.12 Memory requirements.
65
3.13 MTDD for f in Example 3.6.
68
3.14 BDD for f in Example 3.7.
69
3.15 MTBDD for the Walsh spectrum for f in Example 3.7.
72
3.16 Calculation procedure for the Fourier transform.
73
3.17 Calculation of the Fourier spectrum through MTDD.
77
3.18 nvMTDD for the Fourier spectrum for f in Example 3.9.
78
4. I
Shannon tree for n
87
4.2
Subtrees in the Shannon tree for pairs of variables (.1,x2),
4.3
= 4.
( m 2 4 ) .
( a ) Subtree in the Shannon tree for a pair of variables
88
(xz. I C ~ + ~()b, ) QDD non-terminal node.
89
4.4
QDD for n = 4.
89
4.5
Subtrees in the Shannon tree for x l , (x2, xg,xq).
90
4.6
( u ) Subtree in the Shannon tree for ( ~ ~ - ~ , x ~ ,( bx) ~ + ~ ) , Non-terminal node with eight outgoing edges. 90
4.7
Decision tree with nodes with two and eight outgoing edges. 91
4.8
Subtree with the Shannon S2 and QDD non-terminal nodes.
4.9
Decision tree for n outgoing edges.
=
4 with nodes with two and four
91 92
4.10 Decomposition of the domain group GIG= C;.
93
4.11 Decomposition of the domain group G16 = Cz.
94
4.12 Fourier decision diagram for f in Example 4.1.
100
4.13 Complex-valued FNADD for Q 2 .
100
4.14 Decision tree with the Shannon node S2 and FNADD nodes for Q2.
4. I5 One-level multi-terminaldecision tree of f in Example 4.3 on GG.
/Of I03
LIST OF FIGURES
4.16 Two-level multi-terminal decision tree of f in Example 4.3 on G6 = C2 x G3.
4.17 Two-level Fourier decision tree of f in Example 4.3 on the Abelian group G6
= Cz x
G3.
4.18 One-level Fourier decision tree of non-Abelian group S3.
f in Example 4.3 on
4.19 Complex-valued Fourier DT on S3 with MTBDD for Sf(2).
xv
103 I 03 I 03
104
4.20 Complex-valued Fourier DT on S3 with Fourier DT for Sf(2). 104 4.21 Decision tree on G36 = C2 x Cz x C3 x C3.
105
4.22 Decision tree on (236
105
=
C3 x C3 x C2 x C2.
4.23 Complex-valued FNADT on G36 = C,
x C3 x S3.
106
4.24 Complex-valued FNADT on G36 = S3 x S3.
107
4.25 Generalized BDD reduction rules.
108
4.26 Arithmetic transform decision diagram for Example 4.6.
f in the
Ill
4.27 Two-variablefunction generator in Shannon notations.
112
4.28 Two-variablefunction generator.
112
4.29 ACDD representation of f ( z ) for TVFG.
113
4.30 Matrix-valued FNADD representation of f ( z ) for TVFG.
113
4.31 MTBDD representation of Sf(4) for TVFG.
115
4.32 Complex-valued FNADD representationof f for TVFG.
115
4.33 MTBDD representationof f for 2-bit adder.
116
4.34 ACDD representation of f for 2-bit adder.
116
4.35 Fourier DT of f for 2-bit adder.
118
4.36 Matrix-valued FNADD of f for 2-bit adder.
118
4.37 MTBDD representationof Sf(4) for 2-bit adder.
119
4.38 Complex-valued FNADD of f for 2-bit adder.
119
4.39 MTDD for the 3-bit multiplier on G64 = C,
121
x C4 x C4.
4.40 FNADD for the 3-bit multiplier on G64 = Qz x
Q2.
124
xvi
LIST OF FIGURES
4.41 BDD representationfor XOr5.
138
4.42 FNADD for XOr5.
138
4.43 Matrix-valued FNAPDD for XOr5.
139
4.44 FNAPDD for XOr5 with elements of mv nodes.
139
4.45 Complex-valuedFNAPDD for XOr5.
140
4.46 BDD for f in Example 4.1 5.
142
4.47 Matrix-valued FNADD for f in Example 4.15.
142
4.48 Complex-valuedFNADD for f in Example 4.1 5.
143
4.49 Matrix-valued FNAPDD for f in Example 4.15.
143
4.50 FNAPDD for f in Example 4.1 5 with elements of mv nodes. 144 4.51 Complex-valued FNAPDD for f in Example 4.1 5. 4.52 MTDD for f in Example 4.16 on G24 = C2 x C2 x C,
144 x C,.
147
4.53 MTDD for f in Example 4.16 on G24 = C4 x C6.
147
4.54 mvFNAPDD for f in Example 4.1 6.
148
4.55 FNAPDD for f in Example 4.1 6 with elements of mv nodes. 148 4.56 ivFNAPDD for f in Example 4.1 6.
149
4.57 Circuit realization of functions from FNAPDDs.
151
5. I
FFT-like algorithms for n
I 75
5.2
BDT for f in Example 5.1.
I 76
5.3
BDD for F in Example 5.1.
I 76
5.4
Calculation of arithmetic-Haar coefficients through BDT.
I 78
5.5
Calculation of arithmetic-Haar coefficients for f in Example 5.2through BDT.
I 78
5.6
Calculation of arithmetic-Haar coefficients for f in Example 5.2through BDD.
179
6.1
a. The partial Gibbs derivative A: on Z9, b. The partial Gibbs derivative A: on Z,, C . The Gibbs derivative D, on 2,. 196
6.2
The flow-graph of the fast algorithm for calculation of the Gibbs derivative D, on 2,.
= 3.
197
LIST OF FIGURES
xvii
6.3
The partial Gibbs derivative A:’ on G12.
200
6.4
The partial Gibbs derivative A;’ on Glz.
200
6.5
The Gibbs derivative DI2.
201
6.6
The flow graph of the fast algorithm for calculation of the Gibbs derivative D12on G12.
202
6.7
Calculation of the partial Gibbs derivative with respect to x3 for f in Example 3.5.
206
7.I
Linear shift-invariant system.
212
7.2
Classification of linear shift-invariant systems on groups.
213
This Page Intentionally Left Blank
List of Tables
2. I
Group operation of S3.
21
2.2
The unitary irreducible representations of S3 over C.
22
2.3
The group characters of S3 over C.
22
2.4
The set R k 3 ) ( x of ) S3 over C.
23
2.5
Unitary irreducible representations of S3 over GF(11).
23
2.6
The group characters of S, over GF(11).
24
2.7
The set R $ ) ( x ) of S3 over GF(11).
24
2.8
Group operation for the quaternion group Q2.
29
2.9
Irreducible unitary representations of
30
Q2
over C.
2. I0 The discrete Walsh functions wal(i, x).
33
3.1
Summary of differences between the FFT on finite Abelian and finite non-Abelian groups.
42
3.2
Group operation of G2x8.
44
3.3
Unitary irreducible representations of GZx8over C.
45
3.4
The unitary irreducible representations of G6x6over GF(11). 52 xix
xx
CONTENTS
3.5
The group operation of G 3 x 6 .
55
3.6
The unitary irreducible representations of G3x6over C
56
3.7
Complexity of the FFT.
63
3.8
Comparisons of domain groups.
64
3.9
FFT for random functions.
64
3.10 Comparison of the FFT for random functions.
64
3.11 Sizes of MTDDs for f and nvMTDDs for S f .
78
4. I
Truth-table for f in Example 4.1.
99
4.2
Space-time complexity of DTs on GI6.
109
4.3
Complexity of representation of TVFG in terms of the number of levels, non-terminal nodes (ntn), constant nodes (cn) and sizes.
115
Complexities of representing the 2-bit adder in terms of levels, number of non-terminal nodes (ntn), constant nodes (cn), and sizes.
120
Complexities of representing the 2-bit multiplier in terms of the number of levels, non-terminal nodes (ntn), constant nodes (cn) and sizes.
121
Complexities of representingthe 3-bit multiplier in terms of the number of levels, non-terminal nodes (ntn), maximum (max) and minimum (min) number of nodes per level, average number per level (av), constant nodes (cn), sizes (s), and maximum number of edges per level (e).
123
4.7
SBDDs and FNADDs for benchmark functions.
125
4.8
SBDDs and FNADDs for adders and multipliers.
126
4.9
SBDDs, MTBDDs and FNADDs for adders and multipliers.
127
4.4 4.5
4.6
4.10 Sizes of ACDDs, WDDs, CHTDDs and FNADDs.
128
4.11 BDDs and FNADDs for Achilles’ heel functions.
128
4.12 Labels of edges.
137
4.13 Complexity of representationof XOr5 in terms of the number of levels, non-terminal nodes (ntn), constant nodes (cn), and sizes.
141
CONTENTS
xxi
4.14 Complexities of representing f in Example 4.1 5 in terms
of the number of levels, non-terminalnodes (ntn), constant nodes, (cn), and sizes.
146
4.15 Complexities of representing f in Example 4.16 in terms
of the number of levels,non-terminal nodes (ntn), constant nodes (cn), and sizes.
152
5.1
Different spectra for f .
171
5.2
Different expressions for f .
i71
5.3
Number of non-zero coefficients.
I 73
6.1
Group operation of
195
6.2
The group representations of Z9 over C.
195
6.3
The group operation of GI2.
198
6.4
The representationsof G12 over GF(11).
199
8.I
The even and odd parts of the test function.
230
8.2
Fourier spectrum of the test function over C.
230
8.3
The even and odd parts of the test function.
233
8.4
Fourier spectrum of the test function.
233
8.5
Fourier spectrum of the test function in GF(11).
233
29.
This Page Intentionally Left Blank
Acronyms ACDD BDD BDT DD DT DFT FFT FDD FNADD FNADT FNAPDD FNAPDT KDD mvMTDD MTBDD MTBDT MDD MTDD MTDT nvMTDD PKDD QDD SBDD TVFG WDD
Arithmetic transform decision diagram Binary decision diagram Binary decision tree Decision diagram Decision tree Discrete Fourier transform Fast Fourier transform Functional decision diagram Fourier decision diagram on finite non-Abelian groups Fourier decision tree on finite non-Abelian groups Fourier decision diagram on finite non-Abelian groups with preprocessing Fourier decision tree on finite non-Abelian groups with preprocessing Kronecker decision diagram Matrix-valued multi-terminal decision diagram Multi-terminal binary decision diagram Multi-terminal binary decision tree Multiple-place diagram Multi-terminal decision diagram Multi-terminal decision tree Number-valued multi-terminal decision diagram Pseudo-Kronecker decision diagram Quaternary decision diagrams Shared binary decision diagrams Two-variable function generator Walsh decision diagram xxiii
This Page Intentionally Left Blank
1 Signals and Their Mathematical Models Humans interact with their environment using various physical processes. For example, the basic means of communication is through sound waves that are generated in the vocal tract and sensed by the ear. Visual communication is done by electromagnetic radiation that can be sensed by the eye. These as well as physical or mechanical interaction can be viewed as processes where a quantity; air pressure, electromagnetic field, physical bodies or their positions are changing as a function of time. These can in a natural way be interpreted as signals, mathematically described as continuous signals, functions of a real variable, often standing for the time. Natural phenomena, such as sound waves as the term indicate, often possess a periodic structure that can be described and analyzed using the powerful tools of Fourier analysis and other sophisticated concepts of mathematical analysis. 1.1
SYSTEMS
Many phenomena can be processed as continuous systems. A typical example is the room audio system, where the microphone picks up the changes in air pressure and converts it to electric voltage or current variations that are amplified and feed to loudspeakers to produce the same sound, but larger in volume. However, for long it has been known that under some reslrictions a signal can be exactly reproduced from just knowing the values of the signal (function) at discrete but dense enough time instants. With the advent of digital computers this opened the way todigital processing of signal in which many of the limitations of physical electronic components can be avoided. This leads to the case where a (discrete) signal can be treated as afunction on 1
2
SIGNALS AND THNR MATHEMATICAL MODELS
a finite cyclic group instead of the fields of real or complex numbers. The fast methods of computing different representations of discrete signals have enabled digital signal processing that has made possible the modern telecommunication systems and many other wonders of modern life. Digital signal processing has leaped from its traditional area of processing digitally signals that used to be processed analogically to many new applications where the phenomena that are investigated can no more be represented as functions of a real or complex variable. The operations that are possible in the digital world are much more complicated than could be realized analogically. Also, the nature of signals may be very different from the original setting. A typical case is the investigation of logic functions using the same transforms as in digital signal processing and a more extreme example is the processing of the information coded in the DNA sequence using signal processing techniques. When we apply the methods of Fourier analysis to a natural or man-made signals, the measurements or the data generated is represented as functions from a set to another. In principle, we could embed these sets in any mathematical structures, groups, rings, etc., for which the tools of Fourier analysis have been developed. However, to get full benefit from this powerful theory, the underlying structures should reflect at least some of the "true" properties of the signals, just as the cyclic group fits naturally to periodicity. Similarly, the dyadic group and the Walsh transform are able to capture properties of logic functions, and so useful in their representations. When more complex phenomena are studied it may not be possible to fully utilize the power of Fourier type methods if we restrict the domain of the signal to be an Abelian group and in certain fields where non-Abelian groups occur most naturally, such as crystallography, Fourier type methods based on non-Abelian groups are routinely used. In signal processing these methods are still not fully developed, but there are plenty of sporadic examples of the power of the theory. In this book we concentrate in presenting (the theory of) Fourier methods over non-Abelian groups for signal processing and logic design. However, we believe that in due time there will be many more applications in the vast range of topics in which signal processing methods are applied.
1.2 SIGNALS When we observe physical signals, changes in air pressure, electromagnetic field, etc., in analog form using some recording device, the recorded signal is only an approximation of the original due to the errors inherent in any sensors. Likewise, even if we assume that the original physical signal satisfied the requirements of the sampling theorem for exact reproduction, our sampling devices have their inherent errors and, thus, only an approximation of the digital equivalent of the original physical signal can be captured. However, as long as the errors are smaller than the accuracy required for extracting the relevant information the system is fine for practical purposes, and a key clement in engineering practice is to balance the cost and performance of thc overall system.
MATHEMATICAL MODELS OF SIGNALS
3
1.3 MATHEMATICAL MODELS OF SIGNALS
Since the signals are physical processes which spread in space-time they are best modelled by elements of some function spaces. To keep the connection to the real world we usually model the inaccuracies as random quantities (noise) following some probability distribution. Often it is necessary to view the "noisless" signal also as a random process that has a joint distribution with the noise process. There is an extensive literature on fundamentals of stochastic signals and the problems that are related to sampling and estimation of signals [27], [39]. Nevertheless the essence of what is now called the sampling theorem was also known to the earlier mathematicians. The reader is referred to [4], [ 161, [20], [50] for some discussions about the history, different formulations, extensions and generalizations of the sampling theorem. For the sampling theorem on the dyadic group see [35]and later [21]. The extension of the theory to arbitrary locally compact Abelian group is given in [26]. An interpretation of the sampling theorem in Fourier analysis on finite dyadic group is given in [31] and extended to arbitrary finite Abelian and non-Abelian groups in [47] and [50], respectively. In engineering practice the signals are modelled by complex functions of real variables and usually called continuous signals. Those represented by discrete functions, i.e., by functions whose variables are taken from discrete sets, are called discrete signals. These also are divided into two subclasses depending on the range of values the signals can take. The continuous signals of a real amplitude are analog signals, while the discrete signals whose amplitudes belong to some finite sets are digital signals. To take advantage of similar powerful mathematical machinery in dealing with discrete signals, it is necessary to impose some algebraic structure on their domain as well as range. In this setting the signals are defined as functions on groups into fields. Moreover, as it has been shown in [52],the structure of a group is the weakest structure on the domain of a signal that still provides a practically tractable model for most of the signal processing and system theory tasks. We consider discrete signals that are defined on some discrete groups, usually identified with the group of integers 2,or with some group 2, of integers modulo p . In other words, the discrete signals are as functions f : Z -+ X , or f : 2, 4 Z,, where X may be the field of complex numbers C, the field of real numbers R, or the group of integers 2,or some finite field. For example, among Abelian groups, the dyadic group and finite dyadic groups G2,$,n E N , have attained a lot of interest, see for example [ 11, [ 3 ] ,since the Walsh functions [ X I , the group characters of these groups [ 121, and their discrete counterparts, the discrete Walsh functions, take two values +l and -1 and, therefore, the calculation of the Walsh-Fourier spectra can be carried out without multiplication. However, there are real-life signals and systems which are more naturally modelled by functions and, respectively, relations between functions on non-Abelian groups. We will mention some related with electrical engineering practice. Some other examples of such problems are discussed in [ 5 ] .
4
SIGNALS AND THEIR MATHEMATICAL MODELS
As is noted in [23], there are examples in pattern recognition for binary images, which may be considered as a problem of realization of binary matrices, in synthesis of rearrangeable switching networks whose outputs depend on the permutation of input terminals [ 151, [34], in interconnecting telephone lines, etc. An application of non-Abelian groups in linear systems theory is in the approximation of a linear time-invariant system by a system whose input and output are functions defined on non-Abelian groups [24]. See, also [41], [42], 1431, [44], [52]. The application of non-Abelian groups in filtering is discussed in [25], where a general model of a suboptimal Wiener filter over a group is defined. It is shown that, with respect to some criteria, the use of non-Abelian groups may be more advantageous that the use of an Abelian group. For example, in some cases the use of the Fourier transform on various non-Abelian groups results in improving statistical performance of the filter as compared to the DFT. See, also [51]. The fast Fourier transform on finite non-Abelian groups [23], [38], has been widely used in different applications [24], [25]. It may be said that for finite non-Abelian groups the quaternion group has a role equal to that played by the finite dyadic group among Abelian groups [50]. Similarly as with the Walsh transform, i.e., the Fourier transform on finite dyadic groups, the calculation of the Fourier transform on the quaternion does not require the multiplication. Regarding the efficiency of the fast Fourier transform on groups, it has been shown [38] for sample evaluations with different groups that in a multiprocessor environment the use of non-Abelian groups, for example quaternions, may result in many cases in optimal, fastest, performance of the FFT. Moreover, as is shown in [38], the quaternion groups as components of the direct product for the domain group G, in many cases, exhibit optimal performance in the accuracy of calculation. These performances have been estimated taking into consideration the number of arithmetic operations, the number of interprocessor data transfers, and the number of communication lines operating in parallel. In this setting, looking for a suitable finite group structure G which should be imposed on the domain of a discrete signal, it has been shown that the combination of small cyclic groups Cz and quaternion groups in the direct product for G results in groups exhibiting, in most cases, the fastest algorithms for the computation of the Fourier transform. In practical applications, we often refer to topological properties of the algebraic structures we use for mathematical models of signals and systems. The space-time topology of the produced solutions stems from the topology (in the mathematical sense) of thc related algebraic structures. It is intersting [ 141 that some important mathematical notions have been introduced first on more complicated structures, and then extended or transferred to the simpler cases. Differential operators could be mentioned as an example. Concept of Newton-Leibnia derivative, the notion was introduced first for real functions, although the continuum of the real line R is one of the most sophisticated algebraic structures (though the richness of the structure was not fully appreciated at that time). Extension of differentiation to the simple case of finite dyadic groups, was done about two centuries after the first vaguc ideas of differentiators and their applications in estimating the rate of change and the direction of change of a signal [ 131. Moreover, it was motivated by the requirements
MATHEMATICAL MODELS OF SIGNALS
5
of technology related to the interest in various applications of two-valued discrete Walsh functions in transmitting and processing binary coded signals and their realizations within prevalent two-stable state circuits environment. The support set of finite dyadic groups, the n-th order direct product of the basic cyclic group of order 2, produces a binary coding of the sequence of first non-negative integers less than 2n, representing a base for the Boolean topologies often used in system design, including the logic design as a particular example of systems devoted the processing of a special class of signals, the logic signals [ 181, [ 191. The restriction of the order to 2n, and some other inconveniences of the Boolean topology, motivated the recent interest in topologies derived from the binary coding of Fibonacci sequences and their applications [I], 181, [9], [lo], [17],[40]. Use of these structures permits introduction of new Fourier-like transforms [ 11, [7], [ 1 I], enriching the class of transforms appearing in Nature and computers [54], [56], [57]. Various extensions and generalizations of the representations of discrete signals and spectral methods in terms of different systems of not necessarily orthogonal basic functions on Abelian groups [2], and the use of non-Abelian groups in signal processing and related areas, suggested probably first in the introduction of [22], offer new interesting research topics as is shown, for example, in [6], 1281, [29], 1301, 1321, 1331, [36]. For these reasons, we have found it interesting to study Fourier transforms on finite non-Abelian groups, and Fourierlike or generalized discrete Fourier transforms [24] on the direct product of finite not necessarily Abelian groups 1451, 1461, [48], 1491. Some recent results in this area are discussed in [37]. These transforms are defined in terms of basic functions generated as the Kronecker product of unitary irreducible representations of subgroups in the domain groups. This way of generalizing the Fourier transform ensures the existence of fast algorithms for efficient calculation of spectral coefficients in terms of space and time. We call all these transforms Fourier transforms, with the excuse that efficient computation is, in many applications, stronger requirement, than possessing counterparts of all the deep properties of the Fourier transform on R. We also consider the Gibbs derivatives on finite non-Abelian groups, since they extend the notion of differentiation to functions on finite groups through a generalization of the relationship between the NewtonLeibniz derivative and Fourier coefficients in Fourier analysis on R.
6
SIGNALS AND THEIR MATHEMATICAL MODELS
REFERENCES I . Agaian, S., Astola, J., Egiazarian, K., Binary Polynomial Transforms and Nonlinear Digital Filters, Marcel Dekker, New York 1995. 2. Aizenberg, N.N., Aizenberg, I.N., Egiazarian, K., Astola. J., “An introduction to algebraic theory of discrete signals”, Proc. First Int. Workshop on Transforms and Filter Banks, TICSP Series, No. I , Tampere, February 23-25, 1998,70-94. 3. Beauchamp, K.G., Applications of Walsh and Related Functions: With an Introduction to Sequency T h e o q Academic Press, New York, 1984. 4. Butzer, P.L., Splettstosser, W., Stens, R.L., “The sampling theorem and linear prediction in signal analysis”, Jber. d. Dt. Math. -Verein., 90, 1988, 1-70.
5. Chirikjian, G.S., Kyatkin, A.B., Engineering Applications of Noncommutative Hurnionic Analysis, CRC Press, Boca Raton, FL, 2000. 6. Creutzburg, R., Labunets, V.G., Labunets, E.V., “Algebraic foundations of an abstract harmonic analysis of signals and systems”, Proc. First Int. Workshop on Transforms and Filter Banks, TICSP Series, No. I , Tampere, February 23-25, 1998. 30-69. 7. Egiazarian, K., Astola, J., “Discrete orthogonal transforms bassed on Fibonaccitype recursion”, Proc. IEEE Digital Signal Processing Workshop (DSPWS-96), Norway, 1996. 8. Egiazarian, K., Astola, J., “Generalized Fibonacci cubes and trees for DSP applications’’, Proc. ISCAS-96, Atlanta, Georgia, May 1996. 9. Egiazarian. K., Astola, J., “On generalized Fibonacci cubes and unitary transforms”, Applicable Algebra in Engineering, Communication and Computing, Vol. AAECC 8, 1997, 371-377. 10. Egiazarian, K., Gevorkian, D., Astola, J., “Time-varying filter banks and multiresolution transforms based on generalized Fibonacci topology”, Proc. 5th IEEE Int. Workshop on Intelligent Signal Proc. and Communication Systems, Kuala Lumpur, Malaysia, 11-13 Nov. 1997, S16.5.1-S16.5.4.
1 1. EgiaLarian, K., Astola. J., Agaian, S., “Orthogonal transforms based on generalized Fibonacci recursions”, Proc. 2nd Int. Workshop on Spectral Techniques and Filter Banks, Brandenburg, Germany, March 5-7, 1999.
12. Fine, N.J., “On the Walsh functions”, Trans. Amer. Math. Soc., Vol. 65, No. 3, 1949,372-414. 13. Gibbs, J.E., “Walsh spectrometry, a form of spectral analysis well suited to binary digital computation”, NPL DES Rept., Teddington, Middlesex, United Kingdom, July 1967.
MATHEMATICAL MODELS OF SIGNALS
7
14. Gibbs, J.E., Simpson, J., “Differentiation on finite Abelian groups”, National Physical Lab., Teddington, England, DES Rept, No. 14, 1974.
15. Harada, K., “Sequental permutation networks”, IEEE Trans. Computers, V0l.C2 I, 472-479, 1972. 16. Higgins, J.R., “Five short stories about the cardinal series”, Bull Amer: Math. SOC., 12, 1985, 45-89. 17. Hsu, W.-J., “Fibonacci cubes-a new interconnection topology”, IEEE Trans. Parallel Distrib. Syst., Vol. 4, 1993, 3-12. 18. Hurst, S.L., The Logical Processing of Digital Signals, Crane Russak and Edvard Arnold, Basel and Bristol, 1978. 19. Hurst, S.L., Miller, D.M., Muzio, J.C., Spectral Techniques in Digital Logic, Academic Press, New York, 1985. 20. Jerri, A.J., “The Shannon sampling theorem-its various extensions and applications: a tutorial review”, Proc. IEEE, Vol. 65, No. 11, 1977, 1565-1596.
21. Kak, S.C., “Sampling theorem in Walsh-Fourier analysis”, Electron. Lett., Vol. 6, July 1970,447-448. 22. Karpovsky, M.G., Finite Orthogonal Series in the Design of Digital Devices, John Wiley and Sons and JUP, New York and Jerusalem, 1976. 23. Karpovsky, M.G., “Fast Fourier transforms on finite non-Abelian groups”, IEEE Trans. Computers, Vol. C-26, 1028- 1030, 1977. 24. Karpovsky, M.G., Trachtenberg, E.A., “Some optimization problems for convolution systems over finite groups”, Inform. and Control, 34,3,227-247, 1977.
25. Karpovsky, M.G., Trachtenberg, E.A., “Statistical and computational performance of a class of generalized Wiener filters”, IEEE Trans. Inform. Theory, VOI. IT-32, 1986. 26. KluvBnek, I., “Sampling theorem in abstract harmonic analysis”, Mat. -Fiz. &sopis, Sloven. Akad. Vied., 15, 1965,43-48. 27. Kotel’nikov, V.A., “On the carrying capacity of the ’ether’ and wire in telecommunications”, Material for the First All-Union Conference on Questions of Communication, Izd. Red. Upr. Svazy RKKA, Moscow, 1933 (in Russian). 28. Kyatkin, A.B., Chirikjian, G.S., “Algorithms for fast convolutions on motion groups“, App. Comp. Harm. Anal., 9,220-241,2000. 29. Labunets, V.G., Algebraic Theory of Signals and Systems-ComputerSignal Processing, Ural State University Press, Sverdlovsk, 1984 (in Russian).
8
SIGNALS AND THEIR MATHEMATICAL MODELS
30. Labunets, V.C., “Fast Fourier transform on generalized dihedral groups”, in Design Automation Theory and Methods, Institute of Technical Cybernetics of the Belorussian Academy of Sciences Press, Minsk, 1985,46-58 (in Russian). 31. Le Dinh, C.T., Le, P., Goulet, R., “Sampling expansions in discrete and finite Walsh-Fourier analysis”, Proc. 1972 Symp. Applic. Walsh Functions, Washington, DC, 1972,265-27 1. 32. Maslen, D.K., “Efficientcomputation of Fourier transforms on compact groups”, J. Fourier Anal. Appl., 1998, 19-52. 33. Moore, C., Rockmore, D.N., Russell, A., Generic Quantum Fourier Transforms, Tech. Rept. quant-ph/0304064, Quantum Physics e-print Archive, 2003. 34. Opferman, D.C., Tsao, Wu, N.T., “On a class of rearrangeable switching networks”, Bell System Tech. J., SC, 1579-1618, 1971. 35. Pichler, F.R., “Sampling theorem with respect to Walsh-Fourier analysis”, Appendix B in Reports Walsh Functions and Linear System Theory, Elect. Eng. Dept., Univ. of Maryland, College Park, MD, May 1970. 36. Rockmore, D.N., “Some applications of generalized FFTs”, An appendix WID. Healy, in Finkelstein, L., Kantor, W., (eds.), Proc. ofthe DIMACS Workshop on Groups and Computation, June 7- 10, 1995, published 1997, 329-369. 37. Rockmore, D.N., “Recent progress i n applications in group FFTs“, in J. Byrnes, G. Ostheimer, (eds.), Computational Noncomutative Algebra and Applications, NATO Science Series: Mathematics, Physics and Chemistry, Springer, Vol. 1 36, 2004. 38. Roziner, T.D., Karpovsky, M.G., Trachtenberg, L.A., “Fast Fourier transform over finite groups by multiprocessor systems”, IEEE Trans. Acoust., Speech, Signal Processing, Vol. ASSP-38, No. 2, 226-240, 1990. 39. Shannon, C.E., “ A mathematical theory of communication”, Bell Systems Tech. J., Vol. 27, 1948, 379-423. 40. Stakhov, A.P., Algorithmic Measurement Theory, Znanie, Moscow, No. 6, 1979, 64 p. (in Russian). 4 I . StankoviC, R.S., “Linear harmonic translation invariant systems on finite nonAbelian groups”, R. Trappl, Ed., Cybernetics and Systems Research, NorthHolland, Amsterdam, 1986. 42. StankoviC, R.S., “A note on differential operators on finite non-Abelian groups,” Applicable Anal., 2 I , 1986, 3 1-41. 43. StankoviC, R.S., “A note on spectral theory on finitc non-Abelian groups,” 3rd Int. Workshop on Spectral Techniques, Univ. Dortmund, 1988, Oct. 4-6, 1988.
MATHEMATICAL MODELS OF SIGNALS
9
44. StankoviC, R.S., “Fast algorithms for calculation of Gibbs derivatives on finite groups,” Approx. Theory and Its Applications, 7,2, June 199I , I - 19. 45. StankoviC, R.S., Astola, J.T., “Design of decision diagrams with increased functionality of nodes through group theory“, IEICE Trans. Fundamentals, Vol. E86-A, NO. 3, 2003,693-703. 46. StankoviC, R.S., Astola, J.T., Spectral Interpretation of Decision Diagrams, Springer, New York, 2003. 47. StankoviC, R.S., StankoviC, M.S., “Sampling expansions for complex-valued functions on finite Abelian groups”, Automatika, 25, 3-4, 147-150, 1984. 48. StankoviC, R.S., Milenovid, D., JankoviC, D., “Quaternion groups versus dyadic groups in representations and processing of switching fucntions”, Proc. 20th Int. Symp. on Multiple-Valued Logic, Freiburg im Breisgau, Germany, May 20-22, 1999, 19-23. 49. Stankovid, R.S., Moraga, C., Astola, J.T., “From Fourierexpansions toarithmeticHaar expressions on quaternion groups“, Applicable Algebra in Engineering, Commuriication and Computing, Vol. AAECC 12, 2001, 227-253. SO. StankoviC, R.S., StojiC, M.R., Bogdanovid, M.R., Fourier Representation of Signals, NauEna knjiga, Beograd, 1988, (in Serbian).
51. Trachtenberg, E.A., “Systems over finite groups as suboptimal filters: a comparative study”, P.A. Fuhrmann, Ed., Proc. 5th Int. Symp. Math. Theory of Systems and Networks, Springer-Verlag, Beer-Sheva, Israel, 1983, 856-863. 52. Trachtenberg, E.A., “SVD of Frobenius matrices for approximate and multiobjective signal processing tasks”, in E.F. Deprettere, Ed., SVD and Signal Processing, Elsevier North-Holland, AmsterdamNew York, 1988, 33 1-345. 53. Trachtenberg, E.A., “Applications of Fourier Analysis on Groups in Engineering Practices”, in StankoviC, R.S., StqjiC, M.R., StankoviC, M.S., (eds.), Recent Developments in Abstract Harmonic Analysis with Applications in Signal Processing, Nauka, Belgrade and Elektronski fakultet, NiS, 1996, 33 1-403. 54. Trakhtman, A.M., Trakhtman, V.A., Furidanientuls of the Theory of Discrete Signals Defined on Finite Inten~als,Sov. Radio, Moscow, 1975 (in Russian). 55. Walsh, J.L., “A closed set of orthogonal functions”, Amez J. Math., 45, 1923, 5-24.
56. Yaroslavsky, L., Digital Picture Processing, an Introduction, Springer Vcrlag, Heidelberg, 1985. 57. Yaroslavsky, L., “Transforms in Nature and computers: origin, discrete representation, synthesis and fast algorithms”, Pror. First Int. Workshop on Transforms and Filter Banks, TICSP Series, No. I , Tampere, February 23-25, 1998, 3-29.
This Page Intentionally Left Blank
Fourier Analysis on Non-Abelian Groups In this chapter, we present a brief introduction to the group representation theory and harmonic analysis on finite not necessarily Abelian groups. For more details the reader is referred to the voluminous literature on abstract harmonic analysis, e.g. [2], [61, [81, [171, [181. The main idea of abstract harmonic analysis is to decompose a complicated function f into pieces that reflect the structure of the group G on which f is defined. The goal is to make some difficult analysis manageable [ 1 I]. The most widely used groups probably are the real line R and the circle R/27rZ, where Z is the set of integers. In the case of the group G = R/27rZ, a given function f on ( - T , T ) is decomposed as
n=-m
and obviously the same representation extends to the periodic extensions of f . For aperiodic functions on the real line R we have the direct and inverse Fourier transforms
11
12
FOURIER ANALYSIS
From the mathematical topology point of view, the real line R is a locally compact Abelian group and the theory of Fourier analysis can been extended to such groups if the exponential functions used in Fourier analysis on R are replaced by the group representations. Compact groups form a subset of the set of locally compact groups. In the following, we briefly introduce the basic concepts of group representations and Fourier analysis on finite groups, and then discuss in more details Fourier analysis on finite non-Abelian groups. We present the basic definitions and and properties as well as proofs of these properties. The purpose is to give the reader some insight to the representation theory in order to clarify the properties of the structures that we use later.
2.1 REPRESENTATIONS OF GROUPS A representation of a group G on a complex vector space V is a correspondence between the abstract group G and a subgroup of the "concrete" group of linear transformations of V , that is, representation is a homomorphism of G into the group of invertible linear transformations on V. Often the group G and the space V are topologized and the group actions are then assumed to be continuous. In the case of finite groups, the linear transformations are usually identified with matrices. In this setting the following definition of group representations can be introduced. The general linear group GL(n,P ) is the group of ( n x n ) invertible matrices ( n is a natural number) with respect to matrix multiplication, with entries in a field P that can be the field of complex numbers C or a finite field Fq where g is power of a prime p . Thus,
GL(n,P ) = {A
E
Pnx,i det(A) = IA/ # 0)
Definition 2.1 (Group representations) A $finite dinierisional representation of a finite group G is a group homomorphism
R : G 4G L ( n ,C ) .
Notice that for a given z E G, R(z) stands for an ( n x n ) matrix R(z) = [R,.,], i, j = 1,.. . , n. The matrix entries Ri.j of R(z)are continuous functions in discrete topology, analogous to trigonometric functions on the circle or the exponential functions exp(27riln) in terms of which the classical Fourier analysis has been defined. Therefore, they will be used to define the Fourier transform on G. Because any finite group of order N is isomorphic to a subgroup of the symmetric group SAT,the group of permutations of N objects, the elements of which can be explicitly listed as N (unitary) ( N x N ) permutation matrices, there are always nontrivial representations. Every finite-dimensional representation is equivalent (similar) to a representation by unitary matrices [ 8 ] . Thus, if C: is a finite group, every representation is equivalent to a unitary representation. Recall that unitary matrices preserve the inner product defined for two vectors x and y in CT1in a usual manner as ( 5 ,y) = ZTy, where Z
REPRESENTATIONS OF GROUPS
13
is the complex-conjugate of z and T denotes the transposition. Thus, for a unitary matrix Q it holds (Qz,Qy) = ( 2 ,y) for all z, y E C". We denote by U ( n ) the multiplicative group of ( n x n ) unitary matrices, i.e.,
U ( n ) = {Q
E
GL(n,C)lQTQ = I},
where I is the identity matrix. Let R and S be representations of degree n. If there is a subspace W of C" such that SW C W and S(z)w = R ( s ) w for all 2 E G and 20 E W , we say that S is a subrepresentation of R. Clearly, then there is a basis of C" such that R(z) has the form
A representation R(z) is called irreducible if its only subrepresentations are R and 0. Consider the space of complex functions on G, i.e., L = { f l f : G + C} which is a vector space of dimension n = JG/over C . Define the convolution product * tEG
tEG
It is straightforward to check that
* makes L into an algebra over C.
Definition 2.2 Two representations S and R are called equivalent ifthere is a matrix T such that S = T - l R T . 2.1 .I
Complete reducibility
Proposition 2.1 Suppose S : G U ( m )is a subrepresentation of R : G Then R is equivalent to the representution --f
+
U(n).
Proof. Since R is unitary, it leaves the inner product (,) invariant. As S is a subrepresentation of R it follows that there is a subspace U1 5 U such that R(X)Ul c Ul, S ( Z ) U= R(X)U, for all z E G, u E U1. Define the orthogonal complement U? of
U:
=
{u E
U I ( U , U ~= )
IJl
as
o forallul E u,}.
Then, R(x)Uf C U f and V(z) defined as thc restriction of R to lit is a subrepresentation of R. Since U = U1 @ U f , the proposition follows.
14
FOURIER ANALYSIS
By induction R(x) is equivalent to
We present the famous result called Schur lemma that is very useful in establishing certain properties of representations.
Lemma 2.1 (Schur) Let R : G 4U ( n )and S : G I(R,S) bY
U ( m )be two representations arid dejine the space
---f
I(R,S) = {a : Cm 4 C"l@is linear and
Q, 0 S(x) = R(L)o QP,
for all z E G}.
that is, such linear operators that the following diagram is commutative f o r all x
c'"
,c"
@
I
I
~
1 R(-d
S(S)
I
Then. 1. The restriction of R to Kera is a subrepresentation ojR, and the restriction of S to ImQPis a subrepresentation of S.
2. IfR and S are irreducible and 3. lfQ,
E
Q,
#
0, then Q, is an isomorphism.
I(R,R),then there is X E C such that @v = Xu f o r all
QP = XI.
1)
E C", i.e.,
4. I f R and S are both irreducible, then
dirnl(R,S) =
1, if R, S are equivalent, 0 otherwise.
Proof, 1. For
'u
E
Kera, we have QP(R(z)v) = S(z)QP,v= S(z)O= 0. Thus R(x)v E E G and we have that 1) E KerQ,. Let w E ImQ, it follows
KerQPfor all z
REPRESENTATIONS OF GROUPS
15
that w = CPv for some w E Cm. Therefore, S(z)w = S(z)CPu = CPR(z)v = CP(R(z)u)E ImCP implying that Im@ is a subrepresentation of S.
2. If CP # 0, it follows that KerQ # C", and since R is irreducible, KerCP = {O}. Now, CP : C" C" is one to one and since S is irreducible, Im@ = C" and m = n. ---f
3. Let X be an eigenvalue of @ and define VX = { u E CmlCPu = Xu}. Now, VX # (0) and R(z)Vx C VX for all z. Thus the restriction of R to V, is a subrepresentation. Since R irreducible, V, = C". Thus, @ = XI. 4. Consider @, Q # 0, that belong to I(R,S). By 2. they are isomorphisms and it follows that @ satisfies WIS(z)= R(z)@-lfor all z E G.
c "'
@,Y .~ ~
c"
Thus, (@-'Q)R(z)= W1S(s)9 = R(x)(CP-~Q) for all z E G implying that W I Q E I(R,R).By 3. we have = XI, or equivalently 9 = XCP and so dimI(R, S) 5 1. Notice that an irreducible representation of an Abelian group has degree one. This can be seen as follows. Fix an element xo E G and consider the map : C" + C" where CPz,,u = for all x E G R(zo)v. Since G is an Abelian group, Q,,, o R(z) = R(z) o implying that Q X , ) E I(R,R),whence R(zo) = CPxll = X I for some A. Thus, R(Q) is diagonal and R can be irreducible only if n = 1.
ax(,
Definition 2.3 (Character of a representation) The character X R of a representation R is Tr(R)where TrA means the trace of A. Recall that the usual inner product in the space of square integrable functions on G into C , L 2 ( G ) is , in the case of a finite group G XEG
Theorem 2.1 (Schur orthogonality relations)
U ( n ) and S : G + U ( m ) be inequivalent and irreducible 1. Let R : G representations of a group G. Then, (I&,Srs) =
c
XEG
R,, (z)S,,(z)= 0, for all
2,
.?>r, 9 ,
16
FOURIERANALYSIS
where Rj:)(z)denotes the i , j - t h entry of the matrix R ( z ) .
2. In particular
where I- is the degree of R. 3. Let R and S be irreducible representations of G. Then, (XR,XS) =
{ ;GI, RR if if
and S are inequivalent, and S are equivalent.
Proof. I . Consider the matrix
P
=
C R(z)CS-~(X), XEG
where C is an
(72
x m ) matrix to be specified later. Now, P E I ( R , S) since
R(y)P
R(yx)CS(F1)
= X € G
C R(w)CS(u-ly)
=
= PS(y).
UEG
Since R and S are irreducible and inequivalent, we have by claim 2. of Schur lemma that P = 0. Choosing C = E,,, the matrix having 1 in the entry s j and 0 elsewhere, we see that
P,,
R a , ( ~ ) S r s=( ~ 0.)
= XEG
2. Again, let C be any nonzero
71
x n matrix and write
Now, by claim 3. of Schur lemma we have P = XI. Setting C = E,, taking trace on both sides we obtain T r ( P ) = X71, = XI.. since trace is invariant under similarity transform we have
XEG
Thus.
XEG
REPRESENTATIONS OF GROUPS
Setting now C
= Ej,,
17
as above we obtain
XEG
3. Let R and S be inequivalent and irreducible. By definition,
If R and S are equivalent, by invariance of trace and claim 2,
Thus, we can say that if R and S are two representations of G and we have the direct sum representations
where R, are irreducible representations of G. Then,
and
Definition 2.4 (Left regular representation) Let x E G and consider the permutation of elements of G, y x-'y and let L(x) be the corresponding (IGI x IGl) permutation matrix. The representation L(x) is called the left regular representation of G. --f
Lemma 2.2 Let L(x) be the left regular representation of G. Then, xL(x)
{ lGl, 0,
i f x = e, the identity of G, otherwise,
and every irreducible representation R E r is contained in L with the multiplicity r. Thus, if r = (R1, . . . , R k } , then L is similar to direct sum
L
(Ti)Ri
CB . . . CB
(Tk)Rk,
(2.2)
18
FOURIER ANALYSIS
and R,E r
Proof. If z # e, then L(z) is a permutation matrix that fixes no element. Thus, TrR(z) = 0. R(e) is the identity matrix and so TrR(e) = IGI. By ( 2 .I), the multiplicity of R, in L is found by
1
(2.3) follows from (2.2).
2.2 FOURIER TRANSFORM ON FINITE GROUPS Now we can collect the results above and define Fourier Analysis on Finite Groups. Let G be a finite group and r the dual object, i.e., set of all inequivalent irreducible representations of G. The matrix entries R z J ( x and ) as well R Z 3 ( x - l )of, the representations R E r as elements of L 2 ( G )form a complete orthogonal set in L2(G).This statement is due to Weyl and his student Peter [I31 and usually called Peter-Weyl theorem, see for example, [8]. As above, denote by T the degree of R. By normalizing we find that ( T ~ G ~ - ~ R )E ~r,~1 5 ~ iR, j ~ 5 T-~ form , a complete orthonormal system in L2(G).Thus we can expand
f
=
C R E r Cl
and by Parseval relation
lIf1l2=
( f l f )
=
cc
(fi
(T-IGl~l )1'2 R % ~
~l(f,R2J)l2.
RET l
These relations imply that it is reasonable to define a Fourier transform on functions f E L2(G)in such a way that the structure of the non-Abelian group becomes expressed in thc transform, e.g. such that the familiar relations of usual Fourier transform with regard to convolutions hold for convolutions defined by the nonAbelian group structure.
Definition 2.5 (Fourier transform) Let f : G + C be a function, rthe dual object of G, R E
r and denote by r the
FOURIER TRANSFORM ON FINITE GROUPS
19
degree of R. The Fourier transform o f f is dejined by
C f (z)R(z-').
F{f )(R) = Sf(R)= ~-1Gl-l
XEG
Thus, the Fourier transform at representation R is an ( r x T - ) matrix where r is the degree of R . By Peter-Weyl theorem the function f can be recovered by the inverse formula
=
c
Tr(%(R)R(x)).
REl-
We will later give list of properties of Fourier transformations on finite groups. As an example, let us consider the convolution property. Let f ,g E L2(g). Then
rIGl-lF{f
* g H R ) = F { f j(R)F{g}(R),
where the multiplication on the right-hand side is multiplication of ( r x r ) matrices. This can be seen as in the Abelian case. x E G uEG
UEG
xEG
vEG
uEG
Here we will fix the notation that will be used later in the book. Let G be a finite, not necessarily Abelian, group of order g = G. We associate (permanently and bijectively) with each group element a non-negative integer from the set {0,1, . . . , gl}, and 0 is associated with the group identity. Thus, each group element will be identified with the fixed non-negative integer associated with it and with no other element. We assume that G can be represented as a direct product of subgroups G1, . . . , G, of orders g l , . . . , g,, respectively, i.e.,
G = x;=iGrt
g
= nr=2=,gt, 91
5 ~2 5 . . . I gn.
(2.4)
The convention adopted above for the notation of group elements applies to the subgroups G, as well. Provided that the notational bijections of the subgroups and of G are consistently chosen, each x E G can be uniquely represented as n
x
=
c a , ~ i 2, , E G,, i=l
5
t G,
(2.5)
20
FOURIER ANALYSIS
with
where g, is the order of G, and 0 5 x,
< g,, i
= 1 , .. . , n. 0
The group operation o of G can be expressed in terms of the group operations i of the subgroups G,, 2 = 1,.. . , n by 0
~
O
=Y( 5 1
0
1 ~ 15,2 2
~ 2 .,. .
xn
g~yn),
Z,Y E
G , .‘~z,yzE ~
z .
(2.6)
Denote by P the complex field or a finite field and by P ( G ) the space of functions f mapping G into P , i.e., f : G + P. Due to the assumption (2.4) and the relation (2.5),each function f E P ( G ) can be considered as an n-variable function f(zi,...,zn), 5% EG,. We denote by K the cardinality of the dual object r or r ( G ) i.e. number of equivalence classes of irreducible representations of G over P. Each such equivalence class contains just one unitary representation. We shall denote the elements of r in some fixed order by Ro, RI, . . . ,RK-1 and by R,(x) the value of R at z E G. Because we now use subscripts to identify the representations, we, from now on, will use superscripts for the elements of the matrices of the representations e.g. R$”’(x), i , j = 1 , 2, . . . ,r,. Notice that group representations may be used in different orders when this is more appropriate for particular applications. This is the same situation as in the case of Abelian groups , for example, the dyadic group where the group characters are Walsh functions [20], which are used in different orderings [9]. If the group G is representable in the form (2.4), then its unitary irreducible representations can be obtained as Kronecker products of the unitary irreducible representations of subgroups G,, i = 1 , .. . , n. Therefore, the number K of unitary irreducible representations of G is
,=1
where K, is the number of unitary irreducible representations of the subgroup G,. Now, for a given group G of the form (2.4), the index 111 of each unitary irreducible representation R, can be written as: n 711
Cbzwt,
W,
E {0,1,. . . , K ,
-
l}, w E {0,1,.. . , K
-
I},
2=1
with
where K, is the number of unitary irreducible representations of the subgroup G,.
FOURIER TRANSFORM ON FINITE GROUPS
21
Table 2.1 Group operation of SR.
1
2
3
4
5
0 1 1 2 2 0 3 4 4 5 5 3
2 0 1 5 3 4
3 5 4 0 2 1
4 3 5 1 0 2
5 4 3 2 1 0
010
We denote by W, the vector of length K of the values of w, considered as an n-variable discrete function. As noted above, the functions Rk’”(z), w = 0,1,. . . , K - 1, i, j = 1,. . . T , form an orthogonal system in the space P ( G ) and the direct and the inverse Fourier transform of a function f E P ( G )are defined respectively by
c
K-1
f ( ~= )
T~(S~(W)R,(X)).
(2.10)
1u=O
Here and in the sequel we shall assume, without explicitly saying so, that all arithmetical operations are carried out in the field P . Note that if in (2.4), there are two identical non-Abelian subgroups G, = G,, z # j , then the Kronecker product of unitary irreducible representations of the subgroups does not produce the unitary irreducible representations of 6. However, the transform defined in terms of so generated a set of linearly independent functions is still denoted as the (generalized) Fourier transform, with the generalization achieved through the formal application of the decomposition used in FFT-like algorithms, see, for example,
1141.
Example 2.1 Let 5’3 = ( 0 , (132), (123), (12), (13),( 2 3 ) ,0)be the symmetricgroup of permutations of order 3. According to the convention adopted in this book, the group elements of 5’3 will be denoted by 0,1,2,3,4,.5, respectively. Using this notation the group operation of 5’3 is shown in Table 2.1. The unitary irreducible representations of 5’3 over C are given in Table 2.2. The group characters and the set Rk”) of functions representing a basis in the space of complex functions on 5’3 are given in Table 2.3 and Table 2.4.
22
FOURIER ANALYSIS
Table 2.2 The unitary irreducible representations of S, over C .
1 2 3 4
1 1 A 1 I B 1 - 1 c 1 - 1 D 5 1 1 -1 E I=[
:],
A=-2-'[
-TI,
&
B=-2-'[
Table 2.3 The group characters of S, over C.
3
1
-
1
0
4 5
1 1
-
1 1
0 0
1 v 5
-&
1
]'
PROPERTIES OF THE FOURIER TRANSFORM
23
Table 2.4 The set R?>’’(Z) of S, over C.
0 I
2 3 4 5
1
1
1
1
1
- 21 - -1 2
1
-1
1
1 1
-1 -1
1 -2
l
l
-
--1 2
0
0
_ _d3
1
h -h
_ _21
0 d3
-1
--
2
2
0
d
h
2
& --
-2
2
_ -1 2
1
2
1 2
Table 2.5 Unitary irreducible representations of S3 over GF(11).
3
10
c
10
E
Example 2.2 The unitary irreducible representations of 5’:s over the Galois field GF(I1) are shown in Table 2.5. The group characters and the basis R:”’ of Ss over GF (11)are given in Table 2.6 and Table 2.7.
2.3 PROPERTIES OF THE FOURIER TRANSFORM The main properties of the Fourier transform on finite non-Abelian groups are analogous to those of the Fourier transform on Abelian groups, such as for example, the
24
FOURlERANALYSlS
-
Table 2.6 The group characters of S3 over GF(11).
3 4
1 1
1 1
0 0
0 0
5
1
1
0
0
Table 2.7 The set R k J ) ( x of ) S, over GF(11).
0
1
1
I
1
1
2
1
1
3
4 5
1
1 1
1
1 1
0
0 0
1
0
0
1
5
8 3 0 8 3
3 8 0 8 3
5
5
1
5 5
5
10 6 6
PROPERTIES OF THE FOURIER TRANSFORM
25
Walsh transform [ 11, the Vilenkin-Chrestenson transform (41, [ 191, see also [9], [ 121, and in particular to that of the Fourier transform on the real line [3].
Theorem 2.2 The main properties of the Fourier transform on jinite non-Abelian groups are the following: 1. Linearity: Forall a 1 , c ~ E P, f l , f i E C ( G ) ,
a l S f l ( 4+ azSfz(w).
Sa,f,+a,,,(W)=
2. Right group translation: For all r
E
G,
3. Group convolution: For twofunctions f l ,f 2 by (fl
* f2)(7) =
E
C ( G )the convolution is defined
cfl(z)f2(r-14
xEG
4. Relative to the so dejined convolution, the Fourier transform exhibits the following proper9 T w s - l s ( f l * f , ) ( T ) ( w= ) Sfl ( W ) S f * ( W ) .
It should be noted that unlike the Fourier transform on Abeliun groups, u dual statement cannot be formulated since the dual object f does not exhibit a group structure suitable for definition of a convolution of functions on r.
5. Purseval theorem: For all f l , f 2
E
P(G),
where 7 denotes the complex-conjugate o f f , S;, (.) is the conjugate transpose o f s f , ( . ) , i.e., S;,(.) = ( S f , ( . ) ) T . 6. The Wiener-Khinchin theorem: For two functions
f l , f2
E
P(G), the cross-
correlation function is dejined by xEG
The autocorrelation function is the cross-correlation function for f l = f i . Denote by Fr: and FG1the direct and inverse Fourier transform on G, respectively, and by F&the transform such that
26
FOURIER ANALYSIS
With this notation the Wiener-Khinchin theorem on G is dejined by
Bfl,fi
= gF,-
(r,
F G ( f l ) F >( f z ) ) .
Proof. Properties 1, 2 and 3 follow immediately from the definition of the Fourier transform and its inverse. The Parseval theorem can be proved by using the orthogonality and the unitarity of Rw(.), i.e., 9
C
R w E r(G )
WSfl (")q2 ('UI))
ZEG
The Wiener-Khinchin theorem follows from the unitarity of Rw(x),the definition of the Fourier transform and the convolution theorem. 2.4
MATRIX INTERPRETATION OF THE FOURIER TRANSFORM ON FINITE NON-ABELIAN GROUPS
In our further consideration we need the generalized matrix multiplications defined as follows. Definition 2.6 Let A be an ( mx n ) matrix with elements a t j E P, i E { 0,1, . . . , ml}, j E {0,1,. . . , n - l}. Let [ B ]be an ( n x r ) matrix whose elements b J k r j E (0,. . . , n l}, k E {0,1,. . . , T- - l} are ( p x q ) matrices with elements in P. These matrices have the same order within a column, and may have different order in different columns of [B]. We define the product A 0 [B]as the (nz x r ) matrix [ Y ]whose elements Yak, i E { O , 1 , . . . , m - I}, k E { O , 1 , . . . , T - I} are ( p x q ) matrices with elements in P given by ~
7L-
I
The product [B]0 A is defined similarly. Definition 2.7 Let [Z]be an ( mx n ) matrix whose elements z,? i E {0,1, . . . , m I}, j E (0, I,.. . , n - 1) are square matrices of not necessarily mutually equal orders with elements in P. Let [B] be ( n x r ) matrix whose elements bjk, j E
MATRIX INTERPRETATION OF THE FOURIER TRANSFORMON FINITE NON-AWELIAN GROUPS
{0,1,. . . , n - l), k t {0,1,. . . , r - l} are square matrices with elements in P. These matrices have the same order within a column, and may have different orders in different columns of [Z] and [B]. Under the condition that the matrices zij and bjk are of the same order or, if not, that one of them is of the order I , the product of matrices [Z] and [B] is dejined as the ( m x T ) matrix Y = [Z]o [B]whose elements yik E P are given by n-1
,
=O
Definition 2.8 Let [Z] be an ( mx n )matrix whose elements z,], i E (0, 1, . . . ,m l}, j E {0,1,. . . , n - 1) are ( p x y) matrices with elements in P. Let [B] be an ( n x r ) matrix whose elements b,k, j E {0,1,. . . , n - I}, k E {0,1,. . . , r - 1) are ( s x t ) matrices with elements in P. Both z,, and b,k are matrices that have the same order within a column, and may have different orders in different columns. The elementwise Kronecker product of matrices [Z] and [B]is dejined as the ( mx r ) 0 matrix [v]= [z]8 [B]whose elements v& are given by n- I j=O
where @ denotes the ordinary Kronecker product. The generalized matrix multiplication concepts, needed to describe the Fourier transform and its inverse on finite non-Abelian groups, were introduced in [15], [ 161. Although we have never met these definitions, we cannot be sure that they are not to be found in the voluminous literature on matrix calculus. By using the matrix operations thus introduced, the Fourier transform pair defined by (2.9) and (2.10) can be expressed as follows. Let f E P ( G ) be given as a vector f = [ f ( O ) , . . . , f(g - 1)]*. Then its Fourier transform is given by
[ S f ]= g-l[R]-' 0f ,
(2.1 1 )
where [ S f ] = [ S f ( O )., . . , S f ( K- 1)IT,and [RIP'= [b,,] with b,, = rWRB1(q), s E {0,1, . . . ,K - l}, 4 t {0,1,. . . , 9 - 1). The inverse Fourier transform is given by f =
[Rl 0 [ S f ] ,
(2.12)
where (R]= [a,,]with at,= R3(i), i E {0,1,. . . , g - l), j E {0,1,. . . ,K - 1).
27
28
FOURlER ANALYSIS
2.5
FAST FOURIER TRANSFORM ON FINITE NON-ABELIAN GROUPS
The formulation of the fast Fourier transform on a finite decomposable group G of the form (2.4) is based on the consideration of the Fourier transform on G as ndimensional Fourier transform each of them relative to one of the n subgroups Gj of
G. In this setting the Fourier transform can be written as:
Step 2
Step j
During the step j , the variables 7111,. . . ,w - 1 , z3+1,.. . ,z., are fixed and the summation is performed through the variable x3 E G, . The algorithm is probably best explained by an example.
Example 2.3 Let G be the Quaternion (non-Abelian) group Q z of order 8. This group has two generators a and b and the group identity is denoted by e. Ifthe group
FAST FOURIER TRANSFORM ON FINITE NON-ABELIAN GROUPS
29
Table 2.8 Group operation for the quaternion group Q 2 .
0
0
1
2
3
4
5
6
7
0
0
1
2
3
4
5
6
7
1 2
1 2
2 3
3 0
0 1
5 6
6 7
7 4
4 5
3 4
3 4
0 7
1 6
2 5
7 2
4 3
5 0
6 1
5 6
5 6
4 5
7 4
6 7
3 0
2 3
1 2
0 1
7
7
6
5
4
1
0
3
2
operation is written as abstract multiplication, the following relations hold f o r the group generators: b2 = a 2 ,bab-' = a - l , a4 = e. Ifthe following bijection V is chosen
I
x e a V(x)IO I
a2
a3
2
3
b 4
ab 5
a2b 6
a3b 7
then the full group operation is described in Table 2.8. All the irreducible unitary representations over the complex field C are given in Table 2.9. For notation convenience, the entries of column Rq can be denoted as
I, ZA, -I,
ZB, C
-iD
E iD,
where
The dual object of Q 2 is of order 5, since there are five irreducible unitary representations of this group. Four of the representations are ]-dimensional and one is 2-dimensional. The Fourier transform matrix is 1 1 1 1 1 1 1 - 1 1 - 1 1 - 1 1 1 1 1 -1 -1 1 -1 1 -1 -1 1 21 2iB -21 2iA 2E 2iD
1 1 -1 -1 2C
1
1,
-;
-
1 -2iD
30
FOURlER ANALYSIS
Table 2 9 Irreducible unitary representations of QZ over C.
3
I
-1
4
1
1
7l
-1
1
-1
[ -;I 0 1
-1
-l
[Y
-1
'
-:,I
[ o ;]
31
FAST FOURIER TRANSFORM ON FINITE NON-ABELIAN GROUPS
r
-
-
-
1 1 1 1 1 1 1 1 1 -1 1 -1 1 -1 1 -1 1 1 1 1 1 - 1 - 1 - 1 - 1 0 1 -1 1 -1 -1 1 -1 1 21 2iB 2iA 2E 2iD -21 2C -2iD -
0 a 0 0
P
A 0 0
This function f is reconstructed from the spectrum [Sf]as
-
-
-
1 1 1 1 1 1 1 1
1 -1 1 -1 1 -1 1 -1
- (a
-
1
8
1 1 I 1 -1 iA 1 1 - 1 1 -1 iB -1 -1 c -1 1 -iD -1 -1 E -1 1 iD
+ /3 + A) + ( - a + p + + (-a + p + + + (-a + p + + (-a + p + + + (-a + p + + (-a + p + + + (-a + p
( a p A) ( a p A) ( a p A) ( a p A) ( a p A) ( a p A) (a+ p +A)
a+P+A -a+p-A a-p-A
1 8
0-
-(I
-
p +A
2
+ + + +
ia
+ +
+ + + + + +
A) ( a - p - A) (-a - p A) A) ( a - p - A) - (-a - p A) - A) ( a - p - A) (-a - p A) - A) ( a p - A) - (-a p A) - A) - ( a - p - A) - (-a - p + A) - A) - ( a - p - A) ( - a - p A) - A) - ( a - p - A) - (-a - p A) - ( - a + p - A) - ((1 - p - A) (-a - p + A)
-
-
-
-
+ +
+ 00 + a1 + a2 + a3 + + (rq
0.5
+ 06 +a7
32
FOURIER ANALYSIS
-J
Fig. 2.7
FFT on the quaternion group Q2.
where
03
Tr(ISf(4))= 0 Tr(iASf(4))= 4 a = Tr(-ISf(4)) I= 0 = Tr(iBSf(4))= -4a
04
=
05
=
00
01 02
=
=
06 07
=
Tr(CSf(4))= P Tr(-iDSf(4)) = 4X Tr(ESf(4))= -4p Tr(iDSf(4))= -4X.
Direct computation of the Fourier transform requires 64 operations f o r the quaternion group G = Q2. Using the fast Fourier transform it can be computed by using 20 additions. The multiplications by the complex unity i are not considered. The corresponding flow-graph is shown in Figure 2.1. Figure 2.2 shows the flow-graph of the FFT algorithm to calculate the inverse Fourier transform on Q 2 . The quaternion group Q 2 is a group structure which can be considered as the domain of signals defined on a set X8 of eight elements. Using the mapping 2
z = Cxz23--a, i=O
2,
E (0, l},
the same set can be equipped with the structure of the dyadic group of order 8.
FAST FOURIER TRANSFORM ON FINITE NON-ABELIAN GROUPS
33
i
Fig. 2.2 Flow-graph for FFT algorithm for the inverse Fourier transform on Q8 Table 2.70 The discrete Walsh functions wal(i, z).
i , x 0 0
1
1
1
2 3 4 5 6 7
1
2
3
4
5
6
7
1
1
1
1
1
1
1
-1
1
-1
-1
1
-1
1
I
1
-1
-1
1
1
-1
1
-1
-1
I
1
-1
-1
-1 1
I
I
1
1
-1
--1
1
-1
I -1 -1
-1 I
I
1
1
-1
-1
-1
1
-1 -1
-1 I
-1
-1
1
1
-1
I
1
-1
Recall that the dyadic group of order 2" consists of the set of binary 72-tuplex z = ( 2 1 , . . . , z n ) , x1 E {0,1), under the componentwise addition modulo 2. Discrete Walsh functions, the discrete version of Walsh functions are the characters of the dyadic groups [7] and, therefore, form a basis in the space of complex functions on Xg. For n = 3 they are given in Table 2.10. The algorithm for the computation of the Fourier transform on Q 2 can be compared to the algorithm for computation of the Fourier transform on the dyadic group of order 8 shown in Figure 2.3. The number of additions and subtractions to compute the Fourier transform on this group is 24 compared to 20 operations on Q 2 and 4 multiplications by the complex unity. The same algorithm can be used to calculate
34
FOURIER ANALYSIS
Fig. 2.3 FFT on the dyadic group of order 8.
the inverse Walsh transform, since it is a self-inverse transform up to the multiplicative constant 2-". A method for an optimal implementation of the Fourier transform on finite not necessarily Abelian groups in a multiprocessor environment is presented in [ 141. It is clear that the ordering of the subgroups Gj, indicated in (2.4), is not essential from the theoretical point of view. However, this order seems logical in some sense, and in our experience not only notationally convenient, but also useful in some practical considerations. Moreover, as is documented in [ 141, for a fixed set of constituent subgroups G, their ordering as is required in (2.4) minimizes the number of data transfers in the implementation of the fast Fourier transform on G in a multiprocessor environment.
FAST FOURIER TRANSFORM ON FINITE NON-ABELIAN GROUPS
35
REFERENCES 1. Beauchamp, K.G., Applications of Walsh and Related Functions: With an Intro-
duction to Sequency Theory, Academic Press, New York, 1984. 2. Borut, A.O., Raczka, R.T., Theory of Group Representations and Applications, Varszawa. 1977. 3. Butzer, P.L., Nessel, R.J., Fourier-AnalysisandApproximation, Vol. I, Birkhauser Verlag, Base1 and Stuttgart, 1971. 4. Chrestenson, N.E., “A class of generalized functions”, facijic J. Math., 5 , 1955, 17-31.
5. Cooley, J.W., Tukey, C.W., “An algorithm for machine calculation of complex Fourier series”, Math. Computation, Vol. 19, 1965, 297-301. 6. Curtis, C.N., Rainer, I., Representation Theory of Finite Groups and Associative Algebras, Halsted, New York, London, 1962.
7. Fine, N.J., “On the Walsh functions”, Trans. Amer: Math. Soc., Vol. 65, No. 3, 1949, 374-414.
8. Hewitt, E., Ross, K.A.,Abstract Harmonic Analysis, IJI, Springr-Verlag, Berlin, 1963, 1970. 9. Hurst, S.L., Miller, D.M., Muzio, J.C., Spectral Techniques in Digital Logic, Academic Press, Toronto, 1985. 10. Karpovsky, M.G., “Fast Fourier transforms on finite non-Abelian groups”, IEEE Trans. Computers, Vo1.C-26, 1028- 1030, 1977. I 1. Knapp, A.W., “Group representations and harmonic analysis from Euler to Langlands“, Part 2, Notices ofAMS, Vol. 43, No. 5, 537-549. 12. Moraga, C., “On a property of the Chrestenson spectrum”, IEE Proc., Vol. 129, Pt. E, 1982, 217-222. 13. Peter, F., Weyl, H., “Vollstandingkeitderprimitiven Darstellungen einergeschlossen kontinuierlichen Gruppe”, Math. Ann., 97, 1927,737-755. 14. Roziner, T.D., Karpovsky, M.G., Trachtenberg, L.A., “Fast Fourier transform over finite groups by multiprocessor systems”, IEEE Trans. Acoust., Speech, Signal Processing, Vol. ASSP-38, No. 2, 226-240, 1990.
15. StankoviC, R.S., “Matrix interpretation of fast Fourier transform on finite nonAbelian groups”, Res. Rept. in Appl. Math., YU ISSN 0353-6491, Ser. Fourier Analysis, Rept. No. 3, April 1990, 1-31, ISBN 86-8 161 1-03-8.
36
FOURIER ANALYSIS
16. Stankovid, R.S.,“Matrix interpretation of the fast Fourier transforms on finite non-Abelian groups”, Proc. /at. Con& on Signal Processing, Beijing/YO, 22.26.10.1990, Beijing, China, 1187-1190. 17. Stankovid, R.S., Stojid, M.R., Bogdanovid, M.R., Fourier Representation of Signals, NauEna knjiga, Beograd, 1988 (in Serbian). 18. Terras, A., Fourier Analysis on Finite Groups and Applications, Cambridge Uni-
versity Press, Cambridge, UK, 1999. 19. Vilenkin, N.Ya., “A class of orthonormal series”, Izv. Akad. Nauk SSSR, Ser: Mat., Vol. 11, 1947, 363-400.
20. Walsh, J.L., “A closed set of orthogonal functions”, Amer: J. Math., 45, 1923, 5-24.
Matrix Interpretation of the Fast Fourier Transform There exist in each area of science some important concepts representing the corner stones of a whole theory and a corresponding practice, study of which, from different aspects, never goes out of interest. The algorithms for efficient computation of the discrete Fourier transform (DFT), generally known as the fast Fourier transform (FFT),form certainly such a concept in digital signal processing. Recall that the great practical application of the DFT and its importance in signal processing stems from the publication of the famous work by Cooley and Tukey in 1965 [lo], though relatively recently interesting facts were discovered about the history of this algorithm [181. Presently there is variety of algorithms in the quite voluminous literature on FFT, each of them suitable with respect to some a priori assumed criteria of optimality. These criteria are very different and range from the reduction of the time and memory resources needed for the computation to the use of some particular properties of the functions whose DFT is to be determined, or the use of the properties of spectral coefficients which should be calculated. Note for example, the real or pure imaginary functions, the symmetric functions, the functions with a lot of zero values, and similarly for spectral coefficients. For more detail see, for example, [I], [7],[31]. FFT algorithms are extended to be applicable to the calculation of the values of the generalized Fourier transform on finite Abelian groups 121, [8] including the DFT as a particular example. Some other particular examples of this theory, as the Walsh or Vilenkin-Chrestenson transform, found also some important applications in different areas (see, for example, [3], [ 171, [22], [30]). Together with that, FFT algorithms were a base for the formulation of fast algorithms for the implementation of other discrete transforms 37
38
MATRIX INTERPRETATION OF THE FFT
on finite sets. Note as examples the discrete Haar transform, the slant transform, the discrete cosine transform (DCT), etc. More information about these algorithms can be found, for example, in [3] and the references therein. Moreover, the practical applicability of the discrete transforms mentioned above and many others is greatly supported by the existence of the fast algorithms for their implementation. The matrix calculus appears to be a convenient way for representing discrete transforms and for dealing with them from theoretical, practical and educational point of view as well. Among different discrete transforms, the Fourier transform on finite non-Abelian groups is recommended recently as the best choice in some particular applications [24], [25], [26], [54], [55]. As we already noted in Section 2.4, a fast algorithm for the implementation of this transform based on the classical Coley-Tukey FFT [3 11, is formulated in an analytical form in [23]. It seems that an earlier related result on this subject can be found in [ 1I]. Matrix interpretation of FFT on finite non-Abelian groups has been given in [41]. The Fourier transform on finite non-Abelian groups can be studied in a unique setting with the Fourier transform on Abelian groups as well as the classical Fourier transform in the frame of abstract harmonic analysis on groups. However, in the case of Fourier transform on non-Abelian groups there are some important differences which must be appreciated at least in practical applications. According to this fact, our aim in this chapter is twofold. First, we consider a matrix representation of the fast Fourier transform on finite non-Abelian groups introduced in attempting to keep the entire analogy with the Abelian case as much as that is possible in the shape of the derived corresponding fast flow-graphs, and, then, we point out and discuss the main differences of this transform with respect to the fast Fourier transform on finite Abelian groups.
3.1 MATRIX INTERPRETATION OF FFT ON FINITE NON-ABELIAN GROUPS To obtain a fast algorithm for the computation of the Fourier transform on finite nonAbelian groups we use the Good-Thomas method as in the case of the FFT on finite Abelian groups [ 151, [ 161, [52]. It is well known that the definition of the fast Fourier transform (FFT) on an Abelian group G (an algorithm for the efficient computation of the Fourier transform on G) is based upon the factorization of G into the equivalence classes relativc to the subgroups of G. This group theoretical approach to the derivation of fast Fourier transform in the matrix notation can be interpreted as follows. The disclosure of the FFT on a finite Abelian group G of the form (2.4) is based upon the factorization of the Fourier transform matrix into a product of sparse factors. Such a factorization is possible since the Fourier transform matrix on a given Abelian group G of the form (2.4) is representable as the Kronecker product of the Fourier transform matrices on its subgroups G,. As is noted in Section 2.1, the transforms
MATRIX INTERPRETATION OF FFT ON FINITE NON-ABELiAN GROUPS
39
whose basic functions are generated as the Kronecker products of unitary irreducible representations of equal non-Abelian subgroups are also considered as the generaliLed Fourier, or short, Fourier transforms on groups [33]. Each of the factor matrices describes uniquely one step of the fast algorithm implementing the Fourier transform with respect to one particular coordinate z,, i = 1, . . . , n, determined by (2.5). In other words, the i-th step of the FFT can be considered as the restriction of the Fourier transform on the whole group G to the Fourier transform on its i-th subgroup G,. It follows that the i-th factor matrix can be represented as the Kronecker product of the Fourier transformation matrix on G, of order g, at i-th position and the identity matrices of orders gJ, j E (1,. . . , n}\{i}, at all other positions into that Kronecker product. We will extend the same approach to the non-Abelian groups by using the concepts of the generalized matrix multiplications. The matrix [R]in the definition of the Fourier transform on finite non-Abelian groups is the matrix of unitary irreducible representations of G over P. Since G is representable in the form (2.4), the matrix [R]can be generated as the Kronecker product of ( K , x 9,) matrices [R,]of unitary irreducible representations of subgroups G, i E (1, ..., n},i.e.,
a=1
Thanks to the well-known properties of the Kronecker product, the same applies to the matrix [Rl-l, is., for this matrix holds n
[RIP' = @[RZ]-'. o=l
This matrix can be further factorized into the elementwise Kronecker product of n sparse factors [C,], i E { 1,.. . , n } as r1
[Ci] = @[s;],2 = 1,. . . , n , J=1
where (3.1)
and I,,, is an ( u x u ) identity matrix. Each matrix [Ci] describes uniquely one step of the fast Fourier transform performed in n steps. The algorithm is best represented by a flow-graph consisting of nodes connected with branches to which some weights are associated. The matrix representation and the corresponding fast algorithm obtained in such a way is similar to the FFT on finite Abelian groups, but some important differences appear here.
40
MATRIX INTERPRETATION OF THE FF7
As it is known, see, for example [22], the flow-graph of thc i-th step of the FFT on a finite Abelian group G of order g has g input and g output nodes. The output nodes of the (i - 1)-th step are the input nodes for the i-th step. The input nodes of the first step are the input nodes of the algorithm, and respectively, the output nodes of the n-th step are the output nodes of the algorithm except for the normalization with g - ' . In the case of non-Abelian groups, the number of input and output nodes is different for the each step of the algorithm. Only the number of input nodes of the algorithm, i.e., the number of input nodes for the first step of the algorithm equals g. The number of input nodes of the i-th step is glg2.. . gz-lgiKi+l . . . K,, while the number of output nodes is 9192 . . . gi-lKzh'z+l . . . K,. Accordingly, the number of output nodes of the algorithm is equal to K . The positions of non-zero elements of [CZ] determine which nodes will be connected. The weights associated to the branches are equal to the values of these elements. An important difference with respect to the FFT on Abelian groups is that in the case of non-Abelian groups the weights may be matrices, and therefore, according to the notation adopted in this book. will be denoted by bold letters. Denote by k ( i, j ) the branch connecting the output node j with the input node i in the flow-graph of the k-th step of the FFT on a finite group. The weight q k ( i , j )associated to this branch is determined by q'(i,j) = C$-k, where CG-k is the ( 2 , j)-th clement of the matrix [ C T L p The k ] . branches for which the weight is equal to zero, i.e., the branches corresponding to zero elements of [C"-']],do not appear in the flow-graph. Now let us give a brief analysis of the complexity of the algorithm described here. The number of calculations is usually employed as a first approximation to the complexity of an algorithm. Taking into account the g input nodes, the number of nodes in the flow-graph described, and hence, the number of basic operations L ( G ) in the FFT on a finite non-Abelian group G based on this flow-graph, is equal to
where a , = g192. . . g,-lKIK,+l . . . K,. Here by a basic operation in the z-th step of the algorithm we mean the opcration given in a general form by
C = ( A @ F ) +( B @G ) B ' where A, B are matrices of order I I ~ : : T ~while , thc weights F and G are matrices of order T,,, . To obtain the Fourier coefficients S j as they are defined by (2.14), it is
ILLUSTRATIVE EXAMPLES
41
needed to perform the normalization by g-' after the calculation in the n-th step of algorithm is carried out. Recall that in the case of Abelian groups, according to the definition of the unitary irreducible representations, all these matrices reduce to the numbers belonging to P , and hcnce, in that case the Kronecker product and matrix addition in P appearing in the basic operation defined here, reduce to the ordinary multiplication and addition in P. Note that the flow-graph of the fast direct Fourier transform can be transformed into the flow-graph for the implementation of the inverse Fourier transform by a suitable mutual replacement of the input and output nodes, i.e., by considering the output nodes as the input nodes and vice versa. Clearly, the weights in this flow-graph are determined by the elements of the matrix [R]factorized as
[R]= [D1]& [D2]& . . . & [D"], where n
[Dz] = @[EI], i = 1,.. . j=1
with
i
I(K,Xf7),
IE)I
=
p31-
7
I(S,XS,b
3 3 3
< 1, =2>
> 2.
The main differences of the FFT on finite non-Abelian groups relative to the FFT on finite Abelian groups are summarized in Table 3.1. 3.2
ILLUSTRATIVE EXAMPLES
As it is usually the case with problems like those considered here, the algorithm is best explained by some examples.
Example 3.1 Let G 2 x be ~ a given group of order 16. The elements of this group will be denoted according to the convention adopted in this monograph by 0,I,..., 15. The identity of the group is 0, and the group operation is described in Table 3.2. All the unitary irreducible representations over the complex field C are given in Table 3.3. Note that in our notation, according to the definition of the inverse Fourier transfi)rm, the Table 3.3 defines the matrix [R]in (2.10). Note that the group Gax8 dejined in this way can be considered as the directproduct of the cyclic group C2 of order 2 with modulo 2 addition as the group operation, and the quaternion group Q2 of order 8. The group representations of Cz are
[ pt 1,
while these of the group
Q2
are in the left upper ( 5 x 8 ) quadrant in Table 3.3 of group representations. The
MATRIX INTERPRETATION OF THE FFT
42
Table 3.1 Summary of differences between the FFT on finite Abelian and finite nonAbelian groups.
Group G of order g Abelian non-Abelian dual object dual object {X}-the set of group r-the set of unitary irreducible representations characters { X}-has the structure r-does not have a group of a multiplicative group structure Direct Fourier transform S,(w)
= g-lrw
cz::Number f(~)R,(z-'), 9-l Cz!; f ( ~ ) T ( w , ~ ) of spectral coefficients ?(W)
K
;:c
f(.)
=
PI
-
=
Inverse Fourier transform
9
; :c
f(.) = WS,(w)Rw(z)) Fourier transformation matrix
Order of the Fourier transformation matrix
Kx9
Direct Fourier transform in matrix notation
[S,] = 9-l [RIP'C> f f
=
PI
Inverse Fourier transform in matrix notation
O
[Sf1
Number of steps of the fast algorithm
f ^ ( M W , 3;)
[TI gxg f = g-"z]f f = [ZIT 'I 1
Tl
Number of input nodes 1-st step g i-th step glg2 . . . g z - 1 g 2 K t + 1 . . . K , n-th step g1K2 . . . K , Number of output nodes I-st step g i-th step 9192.. . gr-lKiKi+l . . . K , n-th step K Weights in the flow-graph The values of unitary irreducible represen tations
{r&-
(.I}
w E (0,. . ., K
9-1
-
9
Y Y
9 9
9 The values of group characters
{x(.)}
w, 2 E {0,1, . . . , g - 1) 1 } , 2 E ( 0 , . . . , g -- l} Factor of normalization after n-th step Y-'
ILLUSTRATIVE EXAMPLES
43
cardinality of the dual object r of GZx8 is K = 10, and, according to (2.7) can be represented as K = KlKz = 2 .5. Notice that the ordering of the columns in Table 3.3 is the result of the usage of the Kronecker product to build the representations (compare the first 8 rows of the first five colunins in Table 2.9). The transformation matrix of the Fourier transform on Gzx8 can be factorized as follows:
where (3.2) (3.3)
with [&Ir1 as determined in Table 2.9 in Example 2.3. TheJEow-graph of the fast Fourier transform on GIs corresponding to this factorization is shown in Figure 3.1. For simplicity the weights corresponding to the branches of thisJEow-graph are not indicated in this jigure. As we noted above, these weights are determined by the values of the elements of the matrices (3.3) and (3.2)for the first and the second step of the flow-graph, respectively. For example, q1(5,4) = 2iD/8 and q2(3,2 ) = 0. Note that in the example considered above the Fourier transform on the quaternion group Q2 and Cz, is used as the basic module, and hence, is performed directly. In the examples below, we consider the Fourier transform on the symmetric group of permutations of order 3, 5'3, in the same way. However, the fast algorithms for the computation of Fourier transform on these groups can be used here at least for the calculation of the coefficients corresponding to the unitary irreducible representations of order 1 and using the matrix operations for the calculation of the remaining coefficients. Of course, by a complete use of these fast algorithms, and therefore, by avoiding the matrix operations, the algorithms shown here can be easily translated into the ordinary-like FFT algorithms like those used in the case of Abelian groups. A discussion of this approach is given in [33]. The presented matrix interpretation of the FFT on finite non-Abelian groups requires the implementation of matrix operations in some branches of the algorithm. Therefore, the presented algorithms are cfficient providing that the elements of the corresponding matrices are calculated simultaneously on some multiprocessor architectures. In other words, the matrix interpretation assumes that the constituent subgroups G, of G are used as the basic modules, as in the case of Abelian groups, the Fourier transform on G is decomposed into n Fourier transforms on G,, where it is implemented directly. However, if the calculation of matrix-valued Fourier coefficients is decomposed into the calculation of their matrix elements, extending in that way the size of each
44
MATRIX INTERPRETATION OF THE FFT
Table 3.2 Group operation of G x R .
-
0
I
2
3
4
5
6
7
8
9 1 0 1 1 1 2 1 3 1 4 1 5 9 1 0 1 1 1 2 1 3 1 4 1 5
0
0
1
2
3
4
5
6
7
8
1
1
2
3
0
5
6
7
4
9 1 0 1 1
2
2
3
0
1
6
7
4
5 1 0 1 1
8
3
3
0
1
2
7
4
5
6 1 1
9 1 0 1 5 1 2 1 3 1 4
4
4
7
6
5
2
1
0
3 1 2 1 5 1 4 1 3 1 0
5
5
4
7
6
3
2
1 0 1 3 1 2 1 5 1 4 1 1 1 0
6
6
5
4
7
0
3
2
7
7
6
5
4
1
0
3
8
8
9 1 0 1 1 1 2 1 3 1 4 1 5
0
1
2
9
9 1 0 1 1
8 1 3 1 4 1 5 1 2
1
2
3
8
9 1 4 1 5 1 2 1 3
2
3
0
9 1 0 1 5 1 2 1 3 1 4
3
0
1
10
1 0 1 1
I1
11
8
8
8 1 3 1 4 1 5 1 2 9 1 4 1 5 1 2 1 3 9
8 1 1 9
8
1 1 4 1 3 1 2 1 5
8 1 1 1 0
9
2 1 5 1 4 1 3 1 2
9
8 1 1 1 0
3
4
5
6
7
0
5
6
7
4
I
6
7
4
5
2
7
4
5
6
8 1 1
4
7
6
5
2
1
0
3
9
8
5
4
7
6
3
2
1
0
1 4 1 3 1 2 1 5
8 1 1 1 0
9
6
5
4
7
0
3
2
1
1 5 1 4 1 3 1 2
9
8 1 1 1 0
7
6
5
4
1
0
3
2
12
1 2 1 5 1 4 1 3 1 0
13
1 3 1 2 l 5 1 4 1 1 l 0
9
14 15
ILLUSTRATIVE EXAMPLES
Table 3.3 Unitary irreducible representations of G z x sover C.
5 0 1 2 3 4
1
1
1
1
1
-1
1
-1
1
1
1
1
1 1 1
1 -1 -1
-1
5
-1 1 -1
6
I
-1
-1
7
1
8 9 10 11
1
-1 1
1
1 1
1 -1 1 -1 1
1
-1
12 13 14 15 -
1 I 1 I
1
-1 1
-1
1 1 1 -1 -1 -1 -1
-1 1 1
-1 1 -1 -1
1 -1 1
1 iA
1 1 1 -1 - 1 1 I iB 1 -1 c 1 1 ZD 1 -1 E I I iD 1 -1 I -1 -1 ZA -1 I -I -1 -1 iB -1 1 c -1 -1 ZD -1 1 E -1 -1 1 ZD -1
1 1
-1
1
1
1
-1
1
1
1
-1 1 -1
1
ZA -
1 - 1 -1 1 -1 -1 -1 1 -1 -1 -1 1 -1 -1 -1 1
1
I
1
1
ZB
c
-iD E ZD -I
ZB
I iA E
iD C
-iD
45
46
MATRIX INTERPRETATION OF THE FFT
Fig. 3.7 Structure of the flow-graph of the FFT on the group GZx8.
ILLUSTRATIVE EXAMPLES
47
step of the FFT into g, the fast algorithms for the calculation of the FFT on some of the basic constituent subgroups can be applied by taking the advantage of some peculiar properties of group representations of the constituent subgroups. For example, the calculation of Fourier transform on the quaternion group Q 2 can be done without multiplication except the multiplication by the imaginary unit i, see Example 2.3. The statement will be illustrated by the following example.
Example 3.2 Let G32 be the direct product of two cyclic groups C2 of order 2 with modulo 2 addition as the group operation and the quaternion group Q 2 of order 8, i.e., G32 = C2 x Cz x Q2. The Fourier transformation matrix on this group over the complex field can be factorized as
[R]-' = [C'] where
& [C2]6 [C3]
with the matrix [ Q 2 ] described in Example 2.3. The structure of the$ow-graph of the FFT on G32 derived according to this factorization is given in Figure 3.2. For simplicity, the weighting coeficients are not shown at the edges of the graph, and these values are determined by the values of the corresponding matrix elements in the factorization of transform matrices. In this algorithm the Fourier transforms on the cyclic groups C2 and the quaternion group Qz are used as the basic modules, and hence they are performed directly. The C2 is the simplest possible case, but the Fourier transform on Q 2 can be carried out by using the corresponding fast algorithm. If we do not like to work with matrix operations in a FFT flow-graph, and if we use the fast algorithm on Q 2 , the algorithm given in Figure 3.2 can be translated easily into an ordinary FFT like those used in the case of Abelian groups. In that order note that the first four rows of [&Iu1 are identical to some particular Walsh functions on the finite group of order 8, and, hence, some of the values representing the output from the first step of our algorithm can be calculated by using apart of the flow-graph of the corresponding fast Walsh transform. The remaining output values from the first step correspond to the group representation of order 2, and therefore they are (2 x 2) matrices. The elements of these matrices will be calculated independently and for efficiency we will use the fact that each of the matrices I, A, B, C, D, E contains two zero elements. In this way we derive the algorithm for the calculation of Fourier transform on G32 shown in Figure 3.3. This algorithm can be compared with the corresponding algorithm presented in [33]
MATRIX INTERPRETATION OF THE F t 7
ft23)
Fig. 3.2 Structure of the flow-graph for FFT on the group C:32.
ILLUSTRATIVE EXAMPLES
49
Fig. 3.3 Structure of the flow-graph for FFT on the group GS2through a part of fast Walsh transform.
50
MATRIX INTERPRETATION OF THE FF7
Fig. 3.4 Structure of the flow-graph for FFT on the group GS2 using FFT on Qz
ILLUSTRATIVE EXAMPLES
51
shown in Figure 3.4. In this algorithm a different fast algorithm for the calculation of the Fourier transform on the quaternion group described in Example 2.3 is used. Example 3.3 Let S, = (0, (132), (123), (12),(13), (23),0) be thesymmetricgroup ofpermutations of order 3 defined in Example 2.1. Let G 6 x 6 = S, x S, be the direct i.e., G G ~consists G of pairs ( h l ,h2) = g E G e x 6 , where product of S, with h l , h2 E S3. The group operation of G 6 x 6 is specified as follows: for ( h l ,h2) = G 6 x 6 , and (hi,h i ) = g' E G ~ i x 6 ,we have (hl h i , h2 h i ) = g 09' E G6x6. The group operation table is large and we do not write it explicitly. I t can be easily derivedfrom the group operation tablefor S,. The unitary irreducible representations Of G 6 x 6 over the Galois$eld G F ( f 1 )are given in Table 3.4. The Fourier transform matrix on G 6 x 6 can be factorized as
g E
[R6x6]-1 =
[C'] & [ C 2 ] !
where
[C'] = [S3]8 [ I ~ > x ~ I
[c2] = [ I 6 x 6 ] @ [S3] >
with [Sg]-l
=2
[; ;
; l'o
l'o l'o
21 2B 2A 2C 2D 2E
]
with the notation as in Table 2.5 since A = BT,and C, D, E are symmetric matrices. TheJlow-graph of the fast Fourier transform on G 6 x 6 based on thisfactorization is shown in Figure 3.5.
i)
Example 3.4 Let Gax6 be the direct product of the group Z, = ( 0 , l , 2, of integers less than 3 with modulo 3 addition as the group operation, and the symmetric group ofpermutations of order 3 , S, described in Example 2.1. The group Z3 is an Abelian group, and therefore, their representations are given by the matrix of characters [XI as follows: [XI =
[ z; ] 1 1 1 el 1 e2
1
1
za).
where el = -2-'(1 - i f i ) , e 2 = -2-l(1 + The unitary irreducible representations of over C are given in Table 2.2. Hence, G s x 6 consists of pairs ( h l ,h2) = g E G 3 x 6 where hl E Z3 and h2 E S,. The group operation o of G3x6 is specified as follows: for ( h l ,h2) = g E GsxG and ( h i ,hi)= g' E G 3 x 6 we have
s,
i
(hl hi, h2 h i ) = g o g1 E G s x 6 . The group table of G 3 x 6 is given in Table 3.5, while the unitary irreducible representations of G s x 6 are given in Table 3.6.
52
MATRIX INTERPRETATION OF THE FFT
Table 3.4 The unitary irreducible representations of G,jx6 over GF(11).
l l I 10
I A B
c
l l I 1
D I 10 E 1 I I l I A I 1 B I 10 C 1 10 D 1 10 E 1 I I l 1 A 1 l B l 1 0 C 1 1 0 D 1 1 0 E 1 1 I 10 1 A 10 1 B 10 10
6
7
1
8 9
1
10
I I I 1 I 1 1 1 I 1
11
12 13
1
I
14
15
16 17 18
19
20 21 22 23 24 25 26 27 28 29 30
1
1
10
c
I
10
1
10
I
I
1
1
1 1 I
10
D E I A B C D E I A B C D E
1 10
I I
10
31
1
1
32
I
I
33
1
10
74
1
10
35 -
1
10
1
l l I 10 10 10
l 1 1 10 10 10
l 1 l 1 1 1 10 10 10
1 A B
c D E I A
B C D E I A B 0 C 0 D 0 E I01 IOA JOB IOC
10
1 1 I
10
10
101
10
10
10
10
IOD 10E
10
1
IOA IOB 1OC 10D IOE
10
10
101
10
10
10
10
IOA IOB 1OC 10D IOE
10
10
10
1
10
1
10
1
10
1
10
1
R6 I I I I I I
R7 I I I
A A A A A A B B B B B B
A A
c
C C
c C C D D D D D D E E E E E E
I, A, B, C, D, E as in Table 2.5.
101 101 101
A 10A IOA IOA
B B B I OB 1 OB 1OB
C C C
1 oc
1oc 1 OC
D D D 1 OD I OD 1OD E E E 1 OE I OE 1OE
A@A A@B A@C ABD ARE
ILLUSTRATIVE EXAMPLES
Fig. 3.5 Structure of the flow-graph for FFT on the group Gex6.
53
54
MATRIX INTERPRETATION OF THE FFT
The Fourier transformation matrix on Gsx6 can be factorized as
[RIP' = [C'] [C2],
with [Sgl-'
=
'[
1
1 1 1 1 21 2B
1 1 2A
1 -1 2C
1 -1
1 -1
2D 2E
1
The structure of the flow-graph of the fast Fourier transform on according to this factorization is given in Figure 3.6.
G3x6 derived
Example 3.5 Consider the group G24 = C2 x C2 x S3,where C2 is the cyclic group of order 2 and S3 is the symmetric group of permutations of order three. Group representations of C2 over GF(11) are given by the columns of the matrix
This matrix is self-inverse up to a multiplicative constant, and therefore, the Fourier transform on C2 is defined by the transform matrix W-'(l) = 6
[ li]
Table 2.5 shows the group representations of S3 over G F ( 11). The Fourier transform matrix on G ~ isJ defined by
[R24]-l(3)= (W-l(l) 60 W-'(1)
@
[S3]-'(1)) mod (11).
[R24]-'(3)= [C'] & [C2]& [C'], where
ILLUSTRATIVE EXAMPLES
55
Table 3.5 The group operation of G3x6.
__.
0
0
1
2
3
4
5
6
7
8
9
1 0 1 1 1 2 1 3 1 4 1 5 1 6 1 7
0
0
I
2
3
4
5
6
7
8
9
1 0 1 1 1 2 1 3 1 4 1 5 1 6 1 7
1
1
2
0
5
3
4
7
8
6
11
9
2
2
0
1
4
5
3
8
6
7
1 0 1 1
1 0 1 1
3
3
4
5
0
1
2
9
4
4
5
3
2
0
1
I 0 1 1
5
5
3
4
I
2
0
11
6
6
7
8
9
1 0 1 1
7
7
8
6
11
9
8
8
6
7
1011
9
9
9
1 0 1 1
6
7
8
10
1 0 1 1
8
6
1 0 7
8
9
9
9
1 0 1 3 1 4 1 2 1 7 1 5 1 6 9
1 4 1 2 1 3 1 6 1 7 1 5
6
7
8
1 5 1 6 1 7 1 2 1 3 1 4
8
6
7
1 6 1 7 1 5 1 4 1 2 1 3
1 0 7
8
6
1 7 1 5 1 6 1 3 1 4 1 2
9
1 2 1 3 1 4 1 5 1 6 1 7 0
1
2
3
4
5
I
2
1 4 1 2 1 3 1 6 1 7 1 5 2
0
0
5
3
4
I
4
5
3
1 5 1 6 1 7 1 2 1 3 1 4 3
4
5
0
I
2
7
1 6 1 7 1 5 1 4 1 2 1 3 4
5
3
2
0
1
6
1 7 1 5 1 6 1 3 1 4 1 2 5
0
1 0 1 3 1 4 1 2 1 7 1 5 1 6
11
11
3
4
1
2
12
1 2 1 3 1 4 1 5 1 6 1 7 0
1
2
3
4
5
6
7
8
9
1 0 1 1
13
1 3 1 4 1 2 1 7 1 5 1 6
1
2
0
5
3
4
7
8
6
11
9
14
1 4 1 2 1 3 1 6 1 7 1 5 2
0
1
4
5
3
8
6
7
1 0 1 1
9
15
1 5 1 6 1 7 1 2 1 3 1 4 3
4
5
0
1
2
9
1011
6
7
8
16
1 6 1 7 1 5 1 4 1 2 1 3 4
5
3
2
0
1 1 0 1 1
9
8
6
7
17
17
3
4
1
2
0
10
7
8
6
15
16
13
14
12
5
11
9
10
56
MATRIX INTERPRETATION OF THE FFT
Table 3.6 The unitary irreducible representations of G s x 6 over C. X -
0 I
1
1
I
1
1
2 3 4 5 6 7 8 9 10 I1 12 13 14
1 I
A B C D E I A B C D
15
1
16
1
17
1
1
I -1 -1 -1
1
1
1
1
1
I -1 -1
1
1
1 1 I I I
1
E I
1
A
1
B C D
-1
-1 -1 -1
E The notation as in Table 2.2
ILLUSTRATIVE EXAMPLES
Fig. 3.6 Structure of the flow-graph for FFT on the group GSx6.
57
Fig. 3.7 Structure of the flow-graph for FFT on G z s .
COMPLEXITY OF THE FFT
59
Figure 3.7 shows the jow-graph of FFT on G24 derived from this factorization of [R24]-l(3). In thispow-graph, the weights a f the edges are elements of l),and of W(1).
(I&[
As described above, in Example 3.2, with FFT the calculation of one-dimensional Fourier transform on G of the form (2.4) is transferred into successive calculation of 7% Fourier transforms on the constituent subgroups G, of G. Each block in the related flow-graph corresponds to a subgroup G,. In these blocks, calculation of the Fourier transform on G, is performed directly by definition of the transform. Some further reduction of computations can be achieved if it is possible to develop some FFT on the constituent subgroups [54]. The Fourier transform matrix on S3 written in terms of functions Rk'3)(x)shown in Table 2.7, is given by
I s;1(1)=2/
I
1 1 1 1 1 1 1 1 1 10 10 10 2 10 10 2 10 10 0 5 6 0 5 6 0 6 5 2 1 0 1 0
0 9
5 1
6 1
This matrix can be factorized as follows:
SS(1) = 2
1 1 0 0 0 0
0 0 1 0 0 1 1
0 1 0 0 1 0 0
0 1 0 1 0 0 0
0 0 0 0 1 0 1 1
0 0 0 0 0 0
1 1 1 0 0 0 2 1 0 1 0 0 0 0 0 5 6 0 0 0 0 0 0 10 10 10 0 0 0 9 1 1 0 0 0 0 6 5
mod (11).
This factorization produces the FFT on S3 over GF(11). Figure 3.8 shows the flowgraph of this FFTon s,. In this figure, the ordering of elementsof Sf(2) as S f (2)(O.'), S f ( 2 ) ( l > l )Sf(2)(l>'), , Sf(2)(Ol1),makes the graph symmetric. Figure 3.9 shows the FFT on G24 derived by using FFT on ,573. In this flowgraph all the weights are numbers. The output is the vector of number-valued Fourier coefficients Sf(O), S f (l), and elements of the matrix-valued Fourier coefficients S;J)(~), i, j = 0, I, for f .
3.3 COMPLEXITY OF THE FFT In this section, we compare efficiency of the FFT on finite dyadic groups and the quaternion groups. Rationales for this comparison can be found in the following considerations. In some areas as, for example, the switching theory and logic design [36],the application of spectral methods, although theoretically proven very efficient [ 11, [ 3 ] ,
60
MATRIX INTERPRETATION OF THE FFT
Fig. 3.8 Structure of the flow-graph for FFT on S3.
[20], [ 191, 1221, [56],is rather restricted in practice for the exponential complexity of the FFT and related algorithms. In these areas, it is often required to calculate with switching functions of a large number of variables, for example, n > 20 or more. In spectral methods, the switching functions, being dependent on two-valued variables, are naturally considered as functions on the dyadic groups. At the same time, being two-valued functions, the set of all switching functions of n-variables is the dyadic group of order Ci" under the componentwise EXOR performed over truthvectors. These basic features determine complexity of representations, manipulations and calculations with switching functions. If n is large, these features become a restrictive factor for many methods and techniques. For example, determination of coefficients in spectral transform representations, as for example, the Reed-Muller expressions and related Kronecker expressions [ 3 5 ] ,1471, by FFT-like algorithms 141, [ 5 ] , expresses exponential complexity of O(2") and O(722") in terms of space and time. The same applies to arithmetic, Walsh, and other related representations of switching functions [Sl]. Thus, if n is large, for example n > 30, such algorithms are hardly applicable on most of standard computer architectures. In [54], the following question is asked: "The ultimate purpose must be to find out whether this group (the quaternion 622) may be as significant for logic synthesis as it seems to be for filtering and other signal processing tasks". The following considerations, based on 1451 and 1461, are an attempt towards approaching an answer to this question. The switching functions are considered as elements of vector spaces on the dyadic groups and the quaternion groups over C. We assume that in an wvariable switching function f , each triplet of variables zz,x3,zk, where each variable takes values in C,, is replaced by a variable X , on
COMPLEXITY OF THE FFT
Fig. 3.9 Structure of the flow-graph for FFT on
G24
with FFT on &.
61
62
MATRIX INTERPRETATION OF THE FFT
Q2. In that way, the domain group C? for f is replaced by Q:,for n = 3 k , by C2Q$ for n = 3k 1, and by C4Q: for n = 3k 2. C4 is the cyclic group of order 4.
+
+
3.3.1 Complexity of calculations of the FFT We compare the space and time complexity of the FFT on dyadic and quaternion groups by the way of examples of mcnc benchmark functions used in logic design and random generated switching functions. Multiple-output functions f o * f l * . . . * f q - l are represented by the integer equivalents f i determined by the mapping f z = azfi. For hardware limitations, the comparison is restricted to small benchmark functions. For detailed theoretical considerations and further examples we refer to [451. The Walsh transform is the Fourier transform on dyadic groups. Calculation of the Walsh transform by FFT requires more operations than calculation of other spectral transform representations on dyadic groups, as for example, the Reed-Muller, Kronecker, or arithmetic expressions [ 141, since unlike these transforms, the Walsh transform matrix does not contain zero elements. Thus, it is enough to compare the complexity of calculations of Fourier transforms on dyadic and quaternion groups. The derived conclusions holds even stronger for other spectral transform representations on dyadic groups. It is enough to compare the implementations through the FFT. In DDs based calculations, discussed in what follows, we actually implement the FFT. For a given f , efficiency may be achieved thanks to peculiar properties o f f , permitting reduction in DDs. Therefore, for a fair comparison for arbitrary functions, we should refer to calculations through DTs. However, in that case, the number of operations is equal to that in the FFT. Thus, it is again enough to compare just FFTs. Figure 3.10 compares the number of operations in the FFT on dyadic groups of and quaternion groups of order T . It shows that for n > 10, the quaternion order groups require fewer number of operations. For analytical expressions of the number of operations and details of implementation of the FFT on the compared groups, see [451. Table 3.7 shows the number of inputs n, time ( t ) ,and space ( m )required to perform the FFT on dyadic ( d ) and quaternion groups ( 9 ) for some benchmark switching functions of different number of variables n. For rt = 8,9,10,14, we use the domain groups Cg, Cz, Cia, Ci4,and C4Q;, Q;,C2Q2, C d Q $ , respectively. Comparison of the considered groups is given in Table 3.8 in terms of the ratio rt = t d / t q and T, = rnd/rriq of the used time and space. Table 3.9 and Table 3.10 give the same information for random generated functions denoted by fun(n) for 7 1 = 9,10,12.14. Time unit is lmsec = lO-’sec. The space is given in KBytes. Calculations were performed on a 133MHz Pcntium PC with 32MBytes of main memory. Figure 3.1 1 compares time required to perform the FFT on C;r and Q; for different values of T . Figure 3.12 compares the corresponding space requirements. This consideration shows that quaternion groups permit faster implementation of the FFT for large switching functions than dyadic groups, at the price of a comparatively small increase of the required space. For example, for n > 10 for at about ten
cpzt
63
0
1
2
3
r = n/3
4
5
6
Fig. 3.70 Number of operations in FFT.
Table 3.7 Complexity of the FFT.
f adr4 misex 1 rd84 mu14 9sym apex4 clip add mu15 sao2
-
n 8 8 8 8 9 9 9 10
10 10 -
td
rnd
t,
46.39 35.14 46.39 45.27
129.70
99.29
379.80 464.00 466.00 759.00 788.00 1531.00
86.39 86.39 86.39 166.39 164.14 166.39
324.60 336.00 339.00 787.00 833.00 763.00
193.30 196.96 197.38 38 1.89 379.93 375.99
1926.39 78360.00 1926.39 75060.00 1618.28 65320.00
5334.39 5331.19 5 165.25
adr7 14 233360.00 misex3 14 801480.00 mu17 14 123200.00 -
125.80
m9
60.90 33.20 116.10 63.40
73.70 133.90
100.27 89.17 101.39
i
64
MATRIX INTERPRETATION OF THE FFT
Table 3.8 Comparisons of domain groups.
clip
9
0.73
2.28
adr5 mu15 sao2
10
10
1.04 1.06 0.50
2.30 2.31 2.26
adr7 misex3 mu17
14 14 14
0.34 0.09 0.53
2.77 2.77 3.19
10
Table 3.9 FFT for random functions. td
fun(9) 536.00 fun( 10) 2048.00 fun( 12) 39620.00 fun( 14) 949000.00
t
m d
86.39 166.39
,
352.00 916.00 486.39 7 I 10.00 1930.30 94280.00
198.64 390.89 1356.32 5406.07
Table 3.10 Comparison of the FFT for random functions.
fun(l4)
0.10
2.80
COMPLEXITY OF THE FFT
Fig. 3.77
Time requirements.
r 6000 5000
4000
r-
I
2000
I
1000 1-
I
0 L 0
5
10
15
I1
Fig. 3.72 Memory requirements.
20
65
66
MATRIX INTERPRETATION OF THE FFT
times faster implementation of the FFT, three times more space is required. When the number of variables 71 grows, the advantages of quaternion groups increase. It follows that, compared to dyadic groups, the quaternion groups have advantages as domain groups for large switching functions.
3.3.2
Remarks on programming implementation of FFT
It should be noticed that besides the number of operations, the programming implementation of a fast algorithm and the hardware used strongly affects its efficiency in terms of time. This aspect also requires some analysis in determination of a transform suitable for a particular application and taking into account also the hardware provided. For a given spectral transform different factorization formulas can be derived. These formulas are derived from the same given set of factorization rules for different specification of various parameters. The number of different formulas is large and dependent on the transform size. For example, as pointed out in [38], there are possible 258400 and 1.8 x 1013 different formulas for the FFT of the size 25 and 26, respectively. Several different factorizations may require the same number of arithmetic operations, however, the order of implementation of the operations may considerably influence the calculation time extending it from 2 to 10 times. In this respect, a lot of research work has been done in searching for the best implementation formulas for the intended hardware. For instance, methods for determination of the most efficient fast Walsh transform with respect to the implementation time has been considered in [21], [32]. Extensions and generalizations of these approaches to the derivation of algorithms that select the best implementation for a variety of signal processing algorithms have been proposed, for example, in [ 121, [38]. Related work for FFT is given, for example, in 1131, [27], [28]. 3.4
FFT THROUGH DECISION DIAGRAMS
FFT algorithms on either Abelian or non-Abelian groups are based upon the vector representations of discrete functions. It follows from the definition of the FFT and their matrix description that space complexity of the FFT on a decomposable group G of order g approximates O ( g ) .The time complexity is O(ng).Thus, the application of the FFT is restricted to groups of relatively small orders. This restriction can be overcome by performing FFT on the decision diagram representations of discrete functions in the same way as that has been done for various discrete transforms on finite Abelian groups. 3.4.1
Decision diagrams
A decision diagram (DD) is a data structure that can compactly represent discrete functions. It is derived from a Decision trec (DT) by merging decision routes that
FFT THROUGH DECISION DIAGRAMS
67
lead to same final value. In this way they often require much less space that other representations. In this book, we use Multi-terminal decision diagrams (MTDDs) [29] to represent discrete functions on finite not necessarily Abelian groups. MTDDs are a generalization of Multi-terminal binary DDs (MTBDDs) [9]. MTBDDs were defined as a generalization of Binary DDs (BDDs) [6] by permitting complex numbers as the values of constant nodes. Thus, they can be used to represent complex-valued functions on dyadic groups. Multiple-place decision diagrams (MDDs) are a generalization of MTBDDs from dyadic groups to groups which are direct product of cyclic groups of order p , C,, [29], [39], [49], [50]. They are derived by allowing p outgoing edges for each node in the DD. MTDDs are a further generalization of MDDs derived by allowing that nodes at different levels in the MTDD may have different number of outgoing edges. MTDDs are derived by the reduction of Multi-terminal decision trees (MTDTs), which can be introduced through the following considerations. A function f ( z ) = f(z1,.. . , 2,) on G = GI x . . . x G,, where \GI = 91. . . g , is defined by specifying its values on each ( X I , .. . , z,) E G. This can be done in a recursive manner as decision tree as follows. Starting form the root node (level 0) we have gl branches for the choices of values for XI. From each of these g1 nodes (on level I), we have g2 branches for the y2 choices of the values for 2 2 . From each node at level n - 1, we have gn branches to the leaves of the decision tree. Each leave gives the value of the function corresponding to the element ( X I , .. . , 2 , ) of G leading to the leave. Thus derived decision tree is called Multi-terminal decision tree (MTDT), since each node has more than two outgoing branches usually called edges. A Multi-terminal decision diagram (MTDD) for a given f is derived from the MTDT by sharing isomorphic subtrees and deleting the redundant information in the decision tree [37]. Formally, a MTDD can be defined as follows. Definition 3.1 (Multi-terminal Decision Diagram) A MTDD for representation o f f E P ( G ) is a rooted directed acyclic graph D(V,E ) with the node set V consisting of non-terminal nodes and terminal or constant nodes. A non-terminal node is labeled with a variable xzo f f and has gz successors denoted by SUCCI, ( v ) E V with k t G,. A constant node v is labeled with an element from P and has no successors. In a MTDT, the i-th level consists of all the nodes to which the variable xi is assigned. In a MTDD, edges connecting nodes at non successive levels may appear. Cross points are points where such an edge crosses levels in the MTDD. Through cross points, the impact of the deleted nodes from the MTDT is taken into account. The concept of MTDD is explained and illustrated by the following example.
Example 3.6 Consider a function f on GZ4 = C, x Example 3.5. I f f is given by the truth-vector f
=
(22
x SB, described in the
[0,6,2,1,0,0,2,1,1,0,0,0,1,1,1,1,1,1,1,1,1,1,2,2]~,
68
MATRIX INTERPRETATION OF THE FFT
then it can be represented by the MTDD showrz in Figure 3.13. I n thisJigure, xa denotes that the variable xi takes the value j .
I,
f
Fig. 3.13 MTDD for f in Example 3.6.
3.4.2
FFT on finite non-Abelian groups through DDs
From the theory of Good-Thomas FFT, t j e calculation of the Fourier transform on a decomposable group G of order g , can be performed through T L Fourier transforms on the constituent subgroups G, of orders 9%. Calculation of Fourier and Fourier-like transforms through decision diagrams is possible thanks to the recursive structure of decision trees, which is compatible with the recursive structure of Kronecker product representable and related transform matrices. Calculation procedures based on decision diagrams representations of discrete functions have been proposed for various discrete transforms on Abelian groups 191, [49], [50], [53]. The decision diagrams methods on non-Abelian groups are extensions of these for Abelian groups. Therefore, before discussing decision diagram methods on non-Abelian, we first briefly explain by an example the calculation of spectral transform on Abelian groups. The following example illustrates the decision diagram methods for calculation of spectral transforms for the case of the Fourier transform on dyadic group of order 8, i.e., the Walsh transform for n = 3 whose FFT is shown in Figure 2.3. There are several interpretation of such methods, most of them exploiting the recursive structure
FFT THROUGH DECISION DIAGRAMS
69
of the Walsh matrix [9], [29], [47]. In this example, the method will be explained by pointing relationships to the FFT-like algorithm for the Walsh transform.
If
Fig. 3.74 BDD for f in Example 3.7.
Example 3.7 For n
= 3,
the Walsh transform matrix is dejned as 3
1 w(3)-'= 8 @ W(1), 2=1
where the basic Walsh matrix is W
=
[ -;1.
For simplicity, in further cal-
culations the scaling factor will be omitted assuming that dividing by 8 will be pe$ormed after the spectral coeficients are calculated. Due to the properties of the Kronecker product, W(3) can be factorized into the product of three sparse matrices, i.e.,
where c 1
=
c 2
=
c 3
=
W(1) @ 1 2 x 2 63 1 2 x 2 , 12x2 c 3 W(1) 8 1 2 x 2 , 1 2 x 2 63 1 2 x 2 @ W(I),
where I z x 2 is the (2 x 2 ) identity matrix. The matrices C1, Ca, and C3 describe steps in the FFT-like algorithmf o r the Walsh transform in Figure 2.3, pe$orming the Walsh trunsform with respect to the variables
MATRIX /NT€RPRETAT/ONOF THE FFT
70
and 2 3 . In a similar way, matrices C1, C2 and C3 describe processing of nodes in a binary decision tree f o r n = 3 in Figure 3.14. The identity matrix means the identical mapping, thus, there is no processing of levels to which they correspond. Thus, C1 shows that we process the root node by W1, and nodes at the other levels remain unprocessed. Similarly, matrices C2 and C3 show that we process by W(1) nodes at levels to which x2 and x3 are assigned. Therefore, it follows that each node in the decision tree should be processed by W(1). The processing means that we perform calculations determined by W(1) over the subfunctions to which point the outgoing edges of the processed node. For a function represented by the decision tree in Figure 3.14 the calculations of the Walsh spectrum are performed as follows. 21, 2 2
1. The nodes at the level f o r x3 are processed as follows:
fo +f l q3,l
=
[ ] [
w(1)
f2
+f 3
f2 -f3 f4
+f5
],
[ ;] [ ff GG + ] f 4 - f5
43.3
=
W(1)
=
f7
-
f7
2. The nodes f o r 2 2 are processed as
3. The root node f o r 2 1 represents the Walsh spectrum since the calculations at this node are as follows:
FFT THROUGH DECISION DIAGRAMS
+ + f 2 + f 3 + f 4 + f s + f s + f7 + f 3 + f 4 f5 + f7 f a + f l - f 2 f 3 + f4 + f 5 f6 fi f a - f l f 2 + f 3 + f4 - f ; - + f 7 fa +f l +f 2 + f4 fs f7 fo + f2 f 3 f 4 + f5 - + f 7 fa + f 2 - f 3 - f 4 - f s + fc + f i f o fl - + - f4 + f 5 + f7 fa
fl
fa -fl
f2 -
-
-
-
71
-
ffj -
-
-
f6
-
f3 -
f l
-
-
-
f6 f6
fl -
-
f2
f3
f6 -
Thus, this is by definition of the Walsh transform, the Walshspectrum in the Hadaniard order: After the calculations are petformed, we divide the thus calculated spectral coeficients by 8 to determine their proper values. If each step of the calculation is represented by a decision tree, the result is the decision tree for the Walsh spectrum. Figure 3.15 shows this calculation procedure. This example illustrates calculations of the Walsh spectrum over the decision tree, since decision trees has the same structure for all the functions for a given number of variables. However, for a given function f , the same calculations can be performed over the decision diagram for f . The reduction of a decision tree into a decision diagram results in the appearance of cross points, which are processed as nodes whose both outgoing edges point to the same subfunction o f f . That approach can be extended to finite non-Abelian groups thanks to the matrix interpretation of the Fourier transform. Calculation through MTDDs consists in the processing of non-terminal nodes and cross points. A MTDD represents the vector F for f . Each non-terminal node, or the cross point, in the MTDD can be considered as the root node of a subtree of the MTDD. Each subtree represents a subvector in F. In that way, the processing of nodes and cross points in the MTBDD means calculations over subvectors represented by the subtrees rooted at the nodes where arrive the outgoing edges of the processed nodes or the cross points. The calculations are performed through some rules that may be conveniently described by matrices. If this is an identity matrix, the processing reduces to the concatenation of subvectors.
72
MATRIX INTERPRETATION OF THE FFT
Fig. 3.75 MTBDD for the Walsh spectrum for f in Example 3.7
73
FFT THROUGH DECISION DIAGRAMS
b e g i n {procedure} for z = n to 1 for k = 0 to Qz do Determine qz,k by using the rule (3.4). Store [S,] = g l . end{p7-OCedU7-e}
Fig. 3.76 Calculation procedure for the Fourier transform.
Definition 3.2 The operation of concatenation, denoted by 0,over an ordered set of n vectors { A1 , . . . , A,} of order rn is the operation producing a vector S of order nrn consisting of n successive subvectors A1 , . . . ,ArL. Example 3.8 Application of the operation of concatemtion over the set of three vectors A = [ula2a3IT,B = [blb2/13]’, C = [ c l c 2 ~ 3 produces ]~’ the vector D = A oBoC = [nlaa~3b~b2b3ci~2c’3]~.
It follows that the calculation of the Fourier coefficients o f f on a finite decomposable group G of order g given by the decision diagram can be carried out through the following procedure. 3.4.2.1 Calculationprocedure Given 3 function f on the decomposable group G of the form (2.4).
1. Represent f by a MTDD. Denote by Q Lthe number of non-terminal nodes at the 2-th level, i.e., the level corresponding to the variable 2, o f f , in the MTDD.
2. Descent the MTDD in a recursive way level by level starting from the constant nodes at (11 + 1)-st level up to the root node at the level I . 3. For 2 = 12 to 1, process the nodes and cross points by using the rule
(3.4)
easily derived from the matrix factorization of [RIP’. Figure 3.16 shows this procedure expressed in a programming pseudo code. It should be pointed out that there is no matrix computation, and only vector operations are used in the computation of Fourier coefficients through this procedure. This ensures efficiency of the procedure. The vectors are represented by decision diagrams, which permits processing of functions on groups of large orders. The matrix-valued vector determined in the root node q1 is the Fourier spectrum o f f . The procedure is probably best explained through an example by using the matrix notation.
74
MATRIX INTERPRETATION OF THE FFT
Example 3.9 Consider the function f in Example 3.6. The Fourier spectrum for f is calculated as
[ S f ]= [ R 2 4 ] - l ( 3 ) F mod (11). Thus,
where
since the scaling factor 6 . 6 . 2 modulo I 1 reduces to 6. Finally,
In (3.I ) , each matrix [C']describes uniquely one step of thefast Fourier transform performed in n steps. In decision diagrams, the operation in the j-th step of FFT is performed over the nodes and cross points at the j-th level in the decision diagrams. Therefore, the Fourier spectrum o f f in Example 3.6 is calculated through MTDD in Figure 3.13 as follows. Note that in this example, all the calculations are in GF(11). 1. The non-terminal nodes q 3 , 0 , q 3 , 1 , q 3 . 3 and the cross point q 3 , 2 are processed jirst by using the matrix [ S 3 ] - ' in [R]- The input data for the procedure are
'.
the values of constant nodes. In that way, it is determined
q3.1
75
FFT THROUGH DECISION DIAGRAMS
2. The non-terminal nodes mined
q2,0, q2,1
are processed by using W(1). It is deter-
3. The root node is processed by using W(1). I t is determined 2 0
[ 4 ;] 5 3
[;:.I
2 0
06
3
[::] + [j:
5 3
9
10
[ b ;] 9
[ s ;] 2
:o]
3 9
[ a ;] [; :la 9 2
-4
76
MATRIX INTERPRETATION OF THE FFT
Thus,
=
Pfl.
Thus, the Fourier spectrum o f f is equal to the matrix-valued vector determined in q1 and it is equal to that calculated by definition of the Fourier transform. Each step of the calculation can be represented by a MTDD, which results in the MTDD for the Fourier spectrum o f f . Figure 3.17 shows the MTDD f o r the Fourier spectrum generated by using the proposed procedure.
3.4.3
MTDDs for the Fourier spectrum
A MTDD for the Fourier spectrum differs from that representing f in the same way as the FFT algorithms on Abelian groups differ from FFT on non-Abelian groups [41]. Decision diagrams representing the Fourier spectrum off on finite non-Abelian groups are matrix-valued, since the values of constant nodes are the Fourier coefficients. The number of outgoing edges of nodes at the i-th level is determined by the cardinality h’,of the dual object r, of G,. In the MTDD for f , the number of outgoing edges of nodes at the i-th level is equal to the order g, of G,. Efficiency of the MTDD representation o f f depends on the number of different values f can take. In the same way, the efficiency of DDs representation of the Fourier spectrum o f f depends on the number of different Fourier coefficients. In this way, it depends indirectly on the number of different values o f f . In a matrix-valued MTDD (mvMTDD) for S f , the matrices representing values of constant nodes can be represented by MTDDs 1431, in the same way as any matrix can be represented by a MTDD [91. In that way, the number-valued MTDDs (nvMTDDs) are derived [43]. Figure 3.18 shows a nvMTBDD for S f in Example 3.9. It is derived from the mvMTDD in Figure 3.17. The columns of matrices representing values of constant nodes are written as subvectors in a vector, which is then represented by a MTDD. In this figure, to make it clear, some constant nodes are repeated. Thanks to the spectral interpretation of DDs 1481, MTDDs for the Fourier spectrum o f f can be interpreted as Fourier DDs for f [42], [43], 1441. 3.4.4
Complexity of DDs calculation methods
The structure of a FFT is described by the number of levels and by connections of nodes within levels. For a given G, the structure of the FFT is determined by the assumed decomposition into the product of subgroups G, of smaller orders (2.4). For the assumed decomposition, the FFT on G has the same structure, thus the same complexity, for any f . Therefore, in calculation of the Fourier transform through the FFT we do not take into account any peculiar properties a function may have.
FFT THROUGH DECISION DIAGRAMS
61
Fig. 3.17 Calculation of the Fourier spectrum through MTDD.
77
88L E LP 6s I L LP 9zz EO I II 0z 91 I 06
PE L9 I
sz
P6 L8 99 EE PE 82 [I IE
sz
S 103 aa8s
S
' 1
I
U j 3 H l j O N011Vl3WdW3lNI XlHlVW
8L
FFT THROUGH DECISION DIAGRAMS
79
Compactness of a decision diagram for f is based upon deleting isomorphic parts in the decision tree for f . Thus, in calculation of the Fourier transform through DD for f , we do not repeat calculations over identical subvectors in the vector representing values o f f . In that way, unlike FFT, we take into account properties of the processed functions. At each node and the cross point at the i-th level, we perform calculations determined by the Fourier transform on G,. Thus complexity of calculations through DDs is proportional to the number of nodes and cross points in the DD for f , usually denoted as the size of the DD. If a function f has some peculiar properties, as for example symmetry, or decomposability, then the MTDD for f has smaller size, and calculation of the Fourier transform is simpler. In reporting experimental results, the size of a DD is usually considered as the number of non-terminal nodes. It is assumed that for an arbitrary function, the number of cross points is smaller than 30 - 40% of the number of non-terminal nodes [37]. The same as in the FFT, the complexity of calculations depends on the complexity of calculation rules derived from the Fourier transforms on G,. Table 3.1 1 shows the sizes of MTDDs for some mcnc benchmark switching functions and their Fourier transforms. In this table, the multi-output switching functions are represented by Shared binary DDs (SBDDs) [37]. The domain group is of the form C;. For non-Abelian groups, the assumed decompositions are shown. C2 and C4 are the cyclic groups of orders 2 and 4, and Q2 denotes the quaternion group of order 8. The price for the reduced size is the increased number of outgoing edges of nodes. The advantage is the reduced depth of the decision diagrams.
80
MATRIX INTERPRETATION OF THE FFT
REFERENCES 1. Agaian, S., Astola, J., Egiazarian, K., Binary Polynomial Transforms and Nonlinear Digital Filters, Marcel Dekker, New York, 1995. 2. Apple, G., Wintz, P., “Calculation of Fourier transforms on finite Abelian groups”, IEEE Trans., Vol. IT-16, 1970, 233-234.
3. Beauchamp, K.G., Applications of Walsh Series and Related Functions: With an Introduction to Sequency Theory, Academic Press, New York, 1984. 4. Besslich, Ph.W., “Efficient computer method for XOR logic design”. IEE Proc., Part E , Vol. 129, 1982, 15-20.
5 . Besslich, Ph.W., “Spectral processing of switching functions using signal flow transformations”, in Karpovsky, M.G., (Ed.), Spectral Techniques arid Fault Detection. Academic Press, Orlando, FL, 198.5. 6. Bryant, R.E. “Graph-based algorithms for Boolean functions manipulation”, IEEE Trans. Comput., Vol. C-35, No. 8, 1986, 667-69 1. 7. Burrus, C.S., Parks, T.W., DFT/FFT and Convolution Algorithms: Theory arid Implementation, John Wiley, New York, 198.5.
8. Cairns, T.W., “On the fast Fourier transform on finite Abelian groups”, IEEE Trans. on Computers, Vol. C-20, 197 1,569-57 I . 9. Clarke, E.M., McMillan, K.L., Zhao, X., Fujita, M., “Spectral transforms for extremely large Boolean functions”, in Kebschull, U., Schubert, E., Rosentiel, W., Eds., Proc. IFIP WG 10.5 Workshop on Applications ofthe Reed-MuUer E.xpansion in Circuit Design, 16- 17.9.1993, Hamburg, Germany, 86-90. 10. Cooley, J.W., Tukey, J.W., “An algorithm for the machine calculation of complex Fourier series”, Math. Computation, Vol. 19, 196.5, 297-30 I . 1 I . Depcyrot, M., “Fondements algebriques de la transformation de Fourier rapide”,
1El1 1, Centre d’automatique, Ecole des Mines, Fontainebleau, France, 1969.
12. Egner, S., Puschel, M., “Generation of fast discrete signal transforms”, IEEE Trans. or2 Signal Processing, Vol. 49, No. 9, 2001, 1992-2002. 13. Frigo, M., Johnson, S.G., “The fastest Fourier transform i n the West”, MIT-LCSR-728, Massachusetts Institute of Tcchnology, Boston, MA, September 1 1. 1997. 14. Gibbs, J.E., Stankovii., R.S., “Matrix interpretation of Gibbs dcrivativcs on finite groups”, private communication, 1988.
IS. Good, I.J., ‘‘The interaction algorithm and practical Fourier analysis”, J. Roy. Stutist. Soc., scr. B, Vo1.20, 361-372, Addendum, V01.22, 1960, 372-375.
FFT THROUGH DECISION DIAGRAMS
81
16. Good, I.J., “The relationship between two fast Fourier transforms”, IEEE Trans. Comput., Vol. C-20, No. 3, 3 10-317, 1971. 17. Harmuth, H.F., Sequency Theory: Foundations and Applications, Academic Press, New York, 1977. 18. Heideman, M.T., Johnson, D.H., Burrus, C.S., “Gauss and the history of the fast Fourier transform”, Archiv for History of Exact Science, Vol. 34, No. 3, 1985, 265-277. Also in IEEE ASSP Magazine, Vol. I , No. 4, Oct. 1984, 14-21. 19. Hurst, S.L., Miller, D.M., Muzio, J.C., Spectral Techniques in Digital Logic, Academic Press, Toronto, 1985. 20. Hurst, S.L., The Logical Processing ofDigita1 Signals, Crane Russak and Edvard Arnold, Basel and Bristol, 1978. 21. Johanson, J., Puschel, M., “In search of the optimal Walsh-Hadamard transform”, Proc. Int. Conf: on Acoustics, Speech and Signal Processing (ICASSP), 2000, 3347-3350. 22. Karpovsky, M.G., Firiite Orthogonal Series in the Design of Digital Devices, Wiley, New York, 1976. 23. Karpovsky, M.G., “Fast Fourier transforms on finite non-Abelian groups”, IEEE Trans. Comput., Vol. C-26, No. 10, Oct. 1977, 1028-1030. 24. Karpovsky, M.G., Trachtenberg, E.A., “Some optimization problems for convolution systems over finite groups”, Inforni. Control., 34, 1977, 227-247. 25. Karpovsky, M.G., Trachtenberg, E.A., “Fourier transform over finite groups for error detection and error correction in computation channels”, Inform. Control, Vol. 40, NO. 3, 1979, 335-358. 26. Karpovsky, M.G., Trachtenberg, E.A., “Statistical and computational performance of a class of generalized Wiener filters”, IEEE Trans. Information Theory, Vol. IT-32, May 1986, 303-307. 27. Kumhom, P., Design, Optimization arid Implementation of a Universal FFT Processor, Ph.D. thesis, Drexel University, Philadelphia, PA, 2001. 28. Kumhom, P., Johnson, J.R., Nagvajara, P., “Design, optimization, and implementation of a universal FFT processor“,Proc. 13th IEEE Int. ASIC/SOC Conference, 2000, 182-1 86. 29. Miller, D.M., “Spectral transformation of multiple-valued decision diagrams”, Proc. 24th lnt. Symp. on Multiple-Valued Logic, Boston, MA, 22.-25.5.1994, 89-96. 30. Moraga, C., “On some applications of the Chrestenson functions in logic design and data processing”, Mathematics arid Computers in Simulation, 27, 1985,43 1 439.
82
MATRIX INTERPRETATION OF THE FFT
3 1. Nussbaumer, H.J., Fast Fourier Transform and Convolution Algorithms, SpringerVerlag, Berlin, Heidelberg, New York, 1981. 32. Park, N., Prasanna, V., “Cache conscious Walsh-Hadamard transform”, Proc.Int. Conference on Acoustics, Speech, and Signal Processing (ICASSP 2001), Vol. 2, 2001.
33. Roziner, T.D., Karpovsky, M.G., Trachtenberg, L.A., “Fast Fourier transforms over finite groups by multiprocessor system”, IEEE Trans., Vol. ASSP-38, No. 2, 1990, 226-240.
34. Sasao, T., “Representations of logic functions by using EXOR operators”, Proc. IFIP WG 10.5 Workshop on Application of the Reed-Muller Expansion in Circuit Design, 27-29.8.1995, Makuhari, Chiba, Japan, 1 1-20. 35. Sasao, T., “Representations of logic functions by using EXOR operators”, in Sasao, T., Fujita, M., (eds.), Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996, 29-54. 36. Sasao, T., Switching Theory f o r Logic Synthesis, Kluwer Academic Publishers, Boston, 1999. 37. Sasao, T., Fujita, M., Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996. 38. Singer, B.W., Automating the Modeling and Optimization ofthe Peformance of Signal Processing Algorithms, Ph.D. thesis, CMU-CS-01- 156, School of Computer Science, Computer Science Department, Carnegie Melon University, Pittsburgh, PA, December 2001. 39. Srinivasan, A,, Kam, T., Malik, Sh., Brayant, R.K., “Algorithms for discrete function manipulation”, in Proc. In$ Con$ on CAD, 1990, 92-95. 40. Stankovid, R.S., “Matrix interpretation of fast Fourier transform on finite nonAbelian groups”, Res. Rept. in Appl. Math., YU ISSN 0353-6491, Ser. Fourier Analysis, Rept. No. 3, April 1990, 1-3 I , ISBN 86-8 161 1-03-8. 41. StankoviC, R.S., “Matrix interpretation of the fast Fourier transforms on finite non-Abelian groups”, Proc. Int. Con$ on Signal Processing, Beijing/90, 22.26.10.1990, Beijing, China, 1 187-1 190. 42. StankoviC, R.S., “Fourier decision diagrams for optimization of decision diagrams representations of discrete functions”, Proc. Workshop on Post BinaryUltra Large Scale Integration, Santiago de Campostela, Spain, 1996, 8- 12. 43. StankoviC, R.S., “Fourier decision diagrams on finite non-Abelian groups with preprocessing,” Proc. 27th Itit. Symp. on Multiple-Valued Logic, Antigonish, Nova Scotia, Canada, May 1997,281-286.
FFT THROUGH DECISION DIAGRAMS
83
44. StankoviC, R.S., Spectral Transform Decision Diagrams in Simple Questions and Simple Answers, Nauka, Belgrade, 1998. 45. StankoviC, R.S., MilenoviC, D., “Some remarks on calculation complexity of Fourier transforms on finite groups”, Proc. 14th European Meeting on Cybernetics and Systems Research, CSMR’98, April 15- 17, 1998, Vienna, Austria, 59-64. 46. StankoviC, R.S., Milenovie, D., JankoviC, D., “Quaternion groups versus dyadic groups in representations and processing of switching fucntions“, Proc. 20th Int. Symp. on Multiple- Valued Logic, Freiburg im Breisgau, Germany, May 20-22, 1999, 19-23. 47. StankoviC, R.S., StankoviC, M., JankoviC, D., Spectral Transforms in Switching Theory, Definitions and Calculations, Nauka, Belgrade, 1998. 48. Stankovid, R.S., Sasao, T., Moraga, C., “Spectral transform decision diagrams”, Proc. IFIP WG 10.5 Workshop on Application of the Reed-Muller Expansion in Circuit Design, 27-29.8.1995, Makuhari, Chiba, Japan, 46-53. 49. StankoviC, R.S., StankoviC,M., Moraga, C., Sasao, T., “Calculation of VilenkinChrestenson transform coefficients of multiple-valued functions through multipleplace decision diagrams”, Proc. 5th Int. Workshop on Spectral Techniques, March 15-17, 1994, Beijing, China, 107-1 16.
50. Stankovid, R.S., StankoviC, M., Moraga, C., Sasao, T., “Calculation of ReedMuller-Fourier coefficients of multiple-valued functions through multiple-place decision diagrams”, Proc. 24th Int. Symp. on Multiple-valued Logic, Boston, MA, May 22-25, 1994, 82-88. 5 1. StankoviC, R.S., StojiC, M.R., StankoviC, M.S., (eds.), Recent Developments in Abstract Harmonic Analysis with Applications in Signal Processing, Nauka, Belgrade and Elektronski fakultet, NiS, 1996.
52. Thomas, L.H., “Using a computer to solve problems in physics”, in Application of Digital Computers, Ginn, MA, 1963. 53. Thornton, M.A., “Modified Haar transform calculation using digital circuit output probabilities”, Proc. IEEE Int. Car$ on Information, Communications and Signal Processing (1st ICICS), Singapore, Vol. I , September 1997,52-58. 54. Trachtenberg, E.A., “Applications of Fourier Analysis on Groups in Engineering Practices”, in StankoviC, R.S., StojiC, M.R., Stankovit, M.S., (eds.), Recent Developments in Abstract Harmonic Analysis with Applications in Signal Processing, Nauka, Belgrade and Elektronski fakultct, NiS, 1996, 33 1-403. 55. Trachtenberg, E.A., Karpovsky, M.G., “Filtering in a communication channel by Fourier transforms over finite groups”, in M.G. Karpovsky, (ed.), Spectral
Techniques and Fault Detection, , Academic Press, Orlando, FL, 1985.
84
MATRIX INTERPRETATION OF THE FFT
56. Varma, D., Trachtenberg, L.A., “Efficient spectral techniques for logic synthesis”, in T. Sasao, Ed., Logic Synthesis and Optimization, Kluwer Academic Publishers, Boston, 1993, 215-232.
4 Optimization of Decision Diagrams In Section 3.4, MTDTs are used as a data structure which permits efficient calculation of the Fourier transform of functions defined on groups of large orders. In this Chapter, we discuss application of non-Abelian groups in the optimization of decision diagram representations for discrete functions, including switching and multiplevalued (MV) functions as particular examples. The presentation is mainly oriented towards decision diagrams for switching functions, since they are the most often met in practice. However, generalizations to multiple-valued, integer or complex-valued functions is simple. In the theory of decision diagram representations [ 171, [23], decision diagrams are derived by reducing the corresponding decision trees (DTs). The reduction is done by deleting or sharing redundant nodes in a DT depending on the equal or otherwise assignment of the decision variables assigned to the non-terminal nodes in the DT. Reduction is performed through rules suitably formulated depending on the range of represented functions and the nodes used in definition of the decision trees. In that respect, BDD reduction rules, zero-suppressed BDD reduction rules [ 161, and generalized BDD reduction rules [27] are distinguished. There are two general approaches to the optimization of decision diagram representations of switching functions reported in the literature: 1. Given a switching function j , represent it by a binary decision diagram (BDD) [ 3 ] . Do the optimization by changing the meaning of nodes , i.e., by choosing
among the Shannon, positive and negative Davio nodes 1161. In that way BDDs are transferred into Kronecker or pseudo-Kronecker decision diagrams. In the spectral interpretation, that means the optimization by searching among different bases to define the decision tree best suited to the given f 1271. 85
86
OPTIMIZATION OF DECISION DIAGRAMS
2. Given a switching function j , represent it by a BDD. Do the optimization of the representation by coding pairs of binary variables by four-valued variables. In that way, the BDD representation is transferred into a Quaternary decision diagram (QDD) representation [ 181. Further optimization can be done by applying the first approach of optimization to quaternary decision diagrams, which produces quaternary Kronecker and pseudo-Kronecker DDs \ 181. A disadvantage ofthe first method is that it does not permit reduction of levels in the decision tree, except in some special cases when the represented functions possess some particular symmetry properties. That was the motivation for the second approach based on recoding of subsets of variables [ 181. For reasons of the intended practical realizations, the method was restricted to pairs of variables [18], although the extension to any subset of variables can be directly given. Moreover, the method does not relate to the function values and can be generalized to decision diagrams for discrete functions, thanks to the extension of the concept of BDDs 131 into multi-terminal decision diagrams (MTBDDs) [ 5 ] . A drawback of this approach to the optimization of decision diagram representations is that it requires nodes with many outgoing edges. Such nodes change the reduction possibilities in decision trees and usually increase the number of nodes per levels in the derived decision diagrams. The reduction of the number of levels and nodes per level in a decision diagram are thus two opposite requirements. In what follows, we first discuss the problem and give a group-theoretic interpretation of the optimization of decision diagrams by coding subsets of variables. Then we offer a solution of the problem by proposing the use of Fourier decision diagrams on finite non-Abelian groups. 4.1
REDUCTION POSSIBILITIES IN DECISION DIAGRAMS
Definition 3.1 of MTDDs can be further elaborated as follows.
Definition 4.1 A decision diagram (DD)that represents an n-variable discretefilnction f ( X I ,. . . , x,) is a rooted acyclic graph G = (V,E ) with the edge set E and the node set V consisting of non-terminal and constant nodes. A variable x, is assigned to each non-terminal node v E V and is called the decision variable f o r 11. Ifx, takes gz different values, then u has g, outgoing edges labeled with the gz values. All the nodes assigned to the same variable X , form the i-th level in the DD, assuming the same order of variables along the paths from the root node to the constant nodes. Decision diagrams with the same order of variables in all the paths are usually called ordered decision diagrams. Notice that there are also defined Free binary decision diagrams (FBDDs) [2], where the order of variables along different paths may be different. However, the restriction that a variable cannot appear several times in a path is preserved in FBDDs. For a given function f , a decision diagram is derived by the reduction of the corresponding decision tree (DT) that is generated by a recursive application of some
REDUCTION POSSIBILITIES IN DECISION DIAGRAMS
-
87
, -
Fig. 4.1 Shannon tree for n = 4.
expansion rule with respect to each variable z, o f f . The reduction is done by sharing the isomorphic subtrees in the decision trees and by deleting the redundant nodes in the decision trees. A node is redundant if all its outgoing edges point to the same node at a successing level in the decision tree. Therefore, a decision tree is the basic concept in decision diagram representations of discrete functions. In the following discussion, the structure of a decision tree is an important concept [23].
Observation 4.1 The structure of a DT is determined by the number of levels and the number of outgoing edges of non-terminal nodes. Note that the number of outgoing edges of nodes at the i-th level uniquely determines the number of nodes at the ( i 1)-st level in a decision tree. Therefore, we may alternatively say that the structure of the decision tree is determined by the number of levels and the number of nodes per levels. These two statements are equivalent. Binary decision diagrams (BDDs) are derived by the reduction of the Shannon trees (Figure 4.1) by BDD reduction rules [16]. For an n-variable function, the Shannon tree consists of n levels with non-terminal nodes with two outgoing edges. In Figure 4.1, and the following figures, the nodes where the Shannon decomposition or its generalizations are performed, are denoted by S,, with index i denoting the number of outgoing edges. The labels at the edges and the values of constant nodes are not shown in the figures, if their meaning is obvious form the context. Quaternary decision diagrams (QDDs) [ 181 are derived by recoding of pairs of switching variables by four-valued variables, (zi,zj) = X k , xi,z j E ( 0 , I}, Xk E {0,1,2,3). In the Shannon tree (schematic representation in Figure 4.2), that means the replacement of each subtree shown in Figure 4.3(a) by a single node with four outgoing edges (Figure 4.3(b)). In that way, the Shannon tree (Figure 4.1) is replaced by the
+
88
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.2 Subtrees in the Shannon tree for pairs of variables
(21, zz), ( ~ 3 ~ x 4 ) .
quaternary decision tree (QDT) (Figure 4.4) with the number of non-terminal levels reduced to a half compared to the starting Shannon tree (two non-terminal levels in QDT instead of four non-terminal levels in the Shannon tree in this example). However, in QDDs nodes with four outgoing edges are required. The method can be extended to the coding of any subset of variables. For example, the Shannon tree for n = 4 may be decomposed into subtrees corresponding to ~1 and ( 2 2 , 2 3 , 2 4 ) = X 2 , X z E (0,. . . , 7) (Figure 4.5). In this way, the subtree in Figure 4.6(a) is replaced by a single node with eight outgoing edges (Figure 4.6(b)). Thus, the Shannon tree for n = 4 is replaced by the tree in Figure 4.7. In this tree, besides the number of levels, the number of non-terminal nodes is also reduced. However, nodes with eight outgoing edges are required. Alternatively, the subtree in Figure 4.6(u) may be replaced by the subtree in Figure 4.8 consisting of a node with two outgoing edges and two nodes with four outgoing edges. The resulting decision tree is shown in Figure 4.9. A drawback of this method for optimization of Decision diagram representations is the appearance of non-terminal nodes with considerable number of outgoing edges. In that way, by recoding subset of variables, we are changing the structure of the decision tree.
Observation 4.2 In a decision tree, the number of outgoing edges of nodes at the i-th level determine the number of nodes at the ( i + 1)-th level and in that w u y determine the reduction possibilities in the decision tree. By changing the structure of the decision tree, we change the reduction possibilities in the decision tree. To explain this statement, denote by 216 = [ z l ,. . . , Z ( 15)IT,the vector reprcsenting the values of the constant nodes in the Shannon tree for n = 4. A node at the 4 t h level, thus, corresponding to 2 4 , may be deleted if the corresponding subvector of 2 1 6 , Z, = [ Z O , Z I ] ~ .,. . ,Z7 = [ z I ~z15IT , is a constant vector. A node at the 4-th level may be shared if the corresponding subvectors Z,, Z,, i # j , i, j E ( 0 , . . . , 7 }
REDUCTION POSSISlLITIES IN DEClSlON DlAGRAMS
Fig. 4.3 ( u ) Subtree in the Shannon tree for a pair of variables ( z t .2% non-terminal node.
f s4
s4
II Fig. 4.4
QDD for n = 4.
s4
89
( b ) QDD
90
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.5 Subtrees in the Shannon tree for xl,( 2 2 : ~ 3 ~ x 4 ) .
Fig. 4.6 ( a ) Subtree in the Shannon tree for with eight outgoing edges.
( X , - ~ , Z ~ , X 1 )~, (~ b )
Non-terminal node
REDUCTION POSSIBILITIES IN DECISION DIAGRAMS
f
’ s4,
Fig. 4.7 Decision tree with nodes with two and eight outgoing edges.
Fig. 4.8 Subtree with the Shannon S z and QDD non-terminal nodes.
91
92
OPTIMIZATION OF DECISION DIAGRAMS
/ s4
\
’ s4
/ ’
s4
\ s4
Fig. 4.9 Decision tree for n = 4 with nodes with two and four outgoing edges.
are equal. We call this property as the reduction possibility of order 2. It is true generally for the nodes at the n-th level in any decision tree whose structure is equal to the structure of the Shannon tree for any n. Examples are the positive Davio tree and Reed-Muller tree and other related decision trees [ 161. To delete or share a node at the level ( n - 1) in such DT we should consider the reduction possibility of order 4 defined as above, but for subvectors of order 4. Further we should consider the reduction possibility of order 2n-1 for the nodes at the second level in the Shannon tree. In QDDs, the minimum reduction possibility is of order 4 at the n/2-th level, while in the decision tree in Figure 4.9 we consider the reduction possibilities of orders 4 and 8, respectively. In the reduction of non-terminal nodes in the decision tree in Figure 4.7 we should consider the subvectors of order 8 and, therefore, the minimum reduction possibility is of order 8. Thus, the reduction of this DT may be done in only a small number of functions having peculiar symmetry properties. The problem is much more obvious in the case of large functions, when decision trees consist of a considerable number of levels and non-terminal nodes. The problem is also very hard in representations of discrete functions taking a large number of different values, thus, having many constant nodes. In this case, we often need almost coinplete multi-terminal decision tree (MTDT) to represent the function. Therefore, by reducing the number of levels in a decision tree by recoding the subsets of variables, i.e., by using the nodes with the increased number of outgoing edges, we decrease the probability to delete or share a non-terminal node in the decision tree. Looking in the opposite direction, from the top of the decision tree, wc are increasing the number of non-terminal nodes at the second level from 2 in the Shannon trees to 4 in QDDs. Since we are increasing the order of the reduction possibility, we
GROUP-THEORETIC INTERPRETATION OF DD
93
Fig. 4.10 Decompositionof the domain group GI6 = Ci.
are increasing the number of non-terminal nodes per levels in the resulting decision diagram. It follows that the reduction of levels and the number of nodes per levels are contradictory requirements in the optimization of decision diagrams. The same consideration applies to the generalization of decision diagrams for switching functions into decision diagrams for discrete functions. BDDs are generalized into multi-terminal binary decision diagrams (MTBDDs) by allowing integers or complex numbers as the values of constant nodes [ 5 ] . Thus, BDD and MTBDDs are derived from decision trees of the same structure, but differ in the values of constant nodes. In the same way QDDs may be generalized to represent integer or complexvalued functions. The same can be done with other decision diagrams defined through coding of different subsets of variables. In that way we have multi-terminal decision diagrams (MTDDs) consisting of nodes with arbitrary number of outgoing edges. In what follows, we give a group-theoretic interpretation of decision trees for switching and discrete functions to express clearly where the problems in reduction of levels and nodes of a decision diagram originate from and to eventually find a way to solve them.
4.2
GROUP-THEORETIC INTERPRETATION OF DD
In decision diagram representation of an n-variable switching functions f by the Shannon tree, f is considered as a function on the finite dyadic group of order 2n, C; = X;=~C~,where C2 = ( ( 0 , l},cb), and @ denotes the modulo 2 addition (EXOR). Thus, for n = 4 the domain group G of f is G = C; = C, x C2 x CZx C, (Figure 4.10). In QDD representation (Figure. 4.4), through the coding of pairs of variables we change the decomposition of the domain group G of f into the product of two cyclic groups of order four, i.e., G = C, x C, (Figure 4.1 1).
94
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.11
Decomposition of the domain group GI6 = C,".
In the group theoretic approach to discrete functions, the structure of decision trees is determined by the structure of the group G on which a given f is considered. Therefore, the Statement 1 can be reformulated as follows.
Observation 4.3 The structure of a MTDT is determined by the decomposition of the domain group G o f f . The number of levels in the decision tree is equal to the number of subgroups G, into which the domain group G o f f is decomposed as a direct product. The number of outgoing edges of nodes at the i-th level and, thus, the number of nodes at the ( i 1)-st level is determined by the order g L of G,. The number of constant nodes is equal to the order g of G.
+
In a decision tree, by changing the meaning of nodes, we change the order of elements in the vector 2,representing the values of constant nodes. In that way, we may increase the probability of finding identical subvectors in 2, and to reduce the number of nodes in the resulting decision diagram. However, we can not reduce the number of levels, since we do not change the structure of the basic decision tree we are starting from. The same applies to the MTBDDs, Walsh decision diagrams (WDDs), Arithmetic transform decision diagrams (ACDDs) [27] and binary moment decision diagrams (BMDDs) 141, which differ from ACDDs in the reduction rules [27]. As is noted above, the reduction of levels may be achieved by recoding the subset of variables, since i n this way we are changing the structure of decision trees we are starting from. From Statement 4.3, in the group-theoretic approach we have the following statement.
Observation 4.4 The number of levels in a decision diagram may be reduced by using different decompositions of the domain group G off intofewer but larger subgroups. For example, in QDDs we are using the subgroups C4 of order four (Figure 4.1 I ) , instead the subgroups C, of order two (Figure. 4.10). Besides the thus achieved reduction of levcls, further reduction may be done by changing the meaning of nodes. In
GROUP-THEORETIC INTERPRETATION OF DD
95
that way the quaternary pseudo Kronecker decision diagrams are introduced [ 1 81. The same can be done for decision diagrams corresponding to any other decomposition of G. However, as noted above, the use of large subgroups requires nodes with considerable number of outgoing edges. Such nodes decrease the probability for reduction. We recall the spectral interpretation of decision diagrams [27] to look for a solution of this problem caused by the contradictory requirements of reducing both levels and nodes per levels. In a decision diagram, the meaning of nodes determines the labels at the outgoing edges of the node. For example, the labels of outgoing edges of the Shannon, positive Davio and negative Davio nodes are { T t ,xi}, {l,x,} and { l , T i } , respectively [16]. In Walsh transform DDs (WDDs), the labels at the edges are { 1,1 - 2xi} [27]. In QDDs, the labels at the edges are {X:,X,’,X,”,X;}, where X! denotes that X , takes the value j [ 181. In spectral interpretation of decision diagrams, we refer to the labels at outgoing edges of nodes. We say that decision trees are the graphical representations of the Fourier and Fourier-like expansions on finite groups with respect to some particular bases. The basis functions are determined as products of labels at the edges in each path from the root node to the constant nodes in the DT. We speak of the Fourier expansions if we use as basis functions the group characters for Abelian and unitary irreducible representations for non-Abelian groups. For any other basis we say the Fourier-like expansion. In this interpretation, the values of constant nodes in decision trees are the Fourier or Fourier-like spectral coefficients of f . A Fourier or Fourier-like spectrum can be considered as a function on the dual object r of G [lo], [ I 11. The Fourier transform on a finite decomposable group G of order g can be expressed as the composition of n Fourier transforms on the constituent subgroups Gi of orders g,. Therefore, we have the following observation for the Fourier decision diagrams.
Observation 4.5 The number of constant nodes in a Fourier DT is equal to the cardinality K of the dual object r of G. The number of the outgoing-edges of nodes at the i-th level is equal to the cardinality K , of the dual object rt of G,. This observation is not in contradiction with the Observation 4.4. MTBDDs and MTDDs are defined with respect to the identical mapping and, thus, in this case the domain group G and its dual object coincides. Moreover, the dual object r of any Abelian group is isomorphic to G. Therefore, the Statement 4.4 applies to any decision tree on Abelian groups. Statement 4.5 is a generalization involving various decision trees on both Abelian and non-Abelian groups as the domain groups for f . The advantages of decomposition of the domain group G into non-Abelian subgroups are in the following. For non-Abelian groups, it is always K , < g,, since at least one of the unitary irreducible representations R, has the order T, > 1. Therefore, we can define the Fourier decision trees by using the unitary irreducible representations as the basis. In these decision trees, the number of outgoing edges of nodes at the i-th level is equal to K , instead of g , in the case of Fourier and any other
96
OPTIMIZATION OF DECISION DIAGRAMS
decision tree on Abelian groups. Therefore, with Fourier decision diagrams, derived by the reduction of Fourier decision trees, we can use large subgroups G, in the decomposition of G to reduce the number of levels. If these subgroups are non-Abelian groups, DT will consists of nodes with a smaller number of outgoing edges compared to any DT on Abelian groups. Thus, with Fourier decision diagrams on non-Abelian groups 1221, we can reduce both number of levels and nodes per levels in a decision tree.
4.3 4.3.1
FOURIER DECISION DIAGRAMS Fourier decision trees
As it was shown in Section 2.5, the Fourier transform on a finite group G decomposable in the form (2.4) may be considered as the n-dimensional Fourier transform on the constituent subgroups G,. Due to that, the fast calculation algorithm are defined, in which each step performs the Fourier transform on a constituent subgroup Gi of the domain group G. In Example 3.7, it is explained the relationship between steps of FFT-like algorithms and decision diagrams and it is shown how to perform FFTlike algorithms over decision diagrams. This relationship is exploited here in the opposite way, the steps of FFT-like algorithms define decomposition rules for a class of decision diagrams which we call the Fourier decision diagrams. Thus, the Fourier transforms on G'i define mappings used to associate a function to the Fourier decision tree. The labels for edges of these diagrams are determined from the inverse Fourier transforms on Gi as i t follows from spectral interpretation of decision diagrams [25].
Definition 4.2 Denote by S f the Fourier transform coejficients on a finite nonAbelian group G representable in the form (2.4). These coeflcients are determined by using the relation (2.9). From (2.10), a Fourier decision tree on G is de$ned as the decision tree whose nodes a the i-th level represent functions f ( G )=
c
xi E G,, T ~(Sf(w)Ru,;(:&)),
w IE r ,
where r2is the dual object of G,, and ELu,>are the unitary irreducible representations of G,. Clearly, the Fourier decision trees on Abelian groups are derived from these on non-Abelian groups as all the unitary irreducible representations are one-dimensional, i.e., reduce to the group characters. With this definition, Fourier decision trees are decision trees on G in which each path from the root node up to the constant nodes corresponds to a unitary irreducible representation of G. A node at the i-th level has K , outgoing edges denoted by Rw, ( 7 ~ )u~ E F a . In the figures of decision trees we use the short notation w: for Rw,( j ) ,or simply j where the context does not permit misunderstandings. For a given f , the values of constant nodes are the Fourier coefficients of .f. In that way, we perform the inverse Fourier transform in the determination o f f represented
FOURIER DECISION DIAGRAMS
97
by a given Fourier decision tree by using the rule in Definition 4.2 and by following the labels at the edges in the decision tree. The same can be done with any spectral transform decision diagram (STDD) 1271. The following example illustrated Definition 4.2.
Example 4.1 Consider a function f defined on the group G12 = Cz x S3 over the field GF( 11). Table 4.1 defines the values for f . The Fourier transforms on C, and S3 over G F ( 11) are dejined by the transform matrices
W-'(l) = 6
[ 1h ]
1 1 1 1 1 - 1 1 1 10 10 10 1 21 2B 2A 2C 2D 2E [RI-' = 1 1 1 1 1 1 1 1 1 10 10 10 - 21 2B 2A 2C 2D 2E
The inverse transform matrices on C, und
-
1 1 1 1 1 1 10 10 10 1 1 1 21 2B 2A 2C 2D 2E 10 10 10 10 10 10 1 1 1 10 10 10 SC SD 9E 91 SB SA S3 are
1 1 11 1 A 1 1 B IS31 = 1 10 c 1 10 D 1 10 E -. -
OPTIMIZATION OF DECISION DIAGRAMS
98
1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 A 1 B 1 o c 1 0 D 1 0 E 1 I 1 A 1 B 10 c 10 D 10 E
1 l 1 1 1
1 1 1 0
1 1 0
1 1 0 10 10 10 10 10 10 10 1 10 1 10 1
I A B
c
D E 101 10A 10B 1oc 10D 10E
The spectrum of the function f in Table 4. I is 4 9
[Sfl =
:] [::]
2 [ ; 0 2
The spectral coeflcients are the values of constant nodes in a Fourier decision diagram f o r f . The labels at the edges are determined from the matrices W(1) and IS,]. The matrix [R]has a block structure that may be expressed as
This structure is due to the Kroneckerproduct and originates in the decomposition of It determines the structure of the Fourier decision trees for functions on G12. Figure 4.12 shows the Fourier decision tree for f . This tree represents f as
G12.
The tree consists of two subtrees corresponding to the left and the right half of the matrix IR1-l. At each level, the transform with respect to the variable assigned to the level is pegormed. The leji FNA3 node corresponds to the calculation specijied by the submatrix [S,]-' in the upper part of leji hulf of [R]-'. When multiplied by the label 1 at the left outgoing edge of the node 5'2, the calculations are repeated to correspond to the lower part of the left half of [R]-l. Similarly, the right FNA3 node corresponds to the upper sumbatrix [S3]-' in the right half of [RI-'.Notice
FOURIER DECISION DIAGRAMS
Table 4.7
0. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.
99
Truth-table for f in Example 4.1
0s 10
1 0 0 0 0 1 0
11
0
12 13 14 1s
0 1 1 0
00 01 02 03 04
that this part corresponds to x1 = 0, which results in the value 1as the label at the right outgoing edge of the node S,. For x1 = 1, this label has the value 10. When multiplied by the value 10, the right FNA3 node performs calculations specified in the lower part of the right half of [Rl-l.In this way, the Kronecker product structure of [R]-l which determines steps of the FF7: is mapped to the recursive structure of the decision tree. It follows that to calculate values of constant nodes in the Fourier decision tree, we perform the Fourier transform. Conversely, through labels at the edges, we perform the inverse Fourier transform to read values of the function represented as stated in Dejriition 4.2, arid calculations are performed as FFT over the decision tree. As is noted above, the Fourier coefficients corresponding to the group representations with orders T, > 1, w = 1,.. . K - 1, are (rU,x T,) matrices. Therefore, Fourier decision trees on finite non-Abelian groups are matrix-valued decision trees. The matrix-valued coefficients may be also represented by the decision trees by using the method of representation of matrices by the decision diagrams [S]. In that way, we derive the integer-valued or complex-valued Fourier decision trees on finite nonAbelian groups, depending on the values of {Rk’”} that are the values of constant nodes in these decision trees. In that case, thc number of constant nodes in a integer or complex-valued Fourier decision trees on a non-Abelian group is equal to that in Fourier or other decision trees on an Abelian group of the same order. Comparing to the matrix-valued Fourier decision trees, the number of levels in an integer-valued or complex-valued Fourier decision tree is increased to represent the matrix-valued constant nodes. However, some non-terminal nodes still may be saved, since not all the Fourier coefficients are matrix-valued.
100
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.12
Fourier decision diagram for f in Example 4.1.
fig. 4.13 Complex-valued FNADD for Q z .
FOURIER DECISION DIAGRAMS
101
Fig. 4.14 Decision tree with the Shannon node SZ and FNADD nodes for Q 2 .
The non-terminal nodes i n Fourier decision trees are denoted by FA, for Abelian and FNA, for non-Abelian groups. The index i denotes the number of outgoing edges and it is equal to the cardinality K , of the dual object r, of G,. The following example illustrates the effect of the decomposition of the domain group o f f into non-Abelian subgroups. Example 4.2 In a Shannon tree, a subtree of order 8 (Figure 4.6(a) and Figure 4.8) may be replaced by the Fourier subtree on QZ shown in Figure 4.13. Compared to the subtree in Figure 4.6(a), we have two levels and two non terminal nodes, but we keep the reduction possibility of order 4 for S4 nodes and 5 for FNA5 nodes. Compared to the subtree in Figure 4.8, we have a saving of one non-terminal node. Figure 4.14 shows the Fourier decision tree on GIG = Cz x Qz. Compared to the Shannon tree in Figure 4. I, we have reduction of both number of levels and total of nodes. Compared to the QDT in Figure 4.4, we have one level more, but reduced the number of nodes per levels. The number of nodes is increased, but the reduction possibilities and the number of outgoing edges per nodes are much better: The number of nodes per levels equals that in Figure 4.7. Compared to the decision tree in Figure 4.9, we have the same number of levels, but reduced number of nonterminal nodes. Thus, the Fourier DT in Figure 4.14 is a compromise between those in Figure 4.1 and Figure 4.4 and Figure 4.9 with respect to the number of nodes and number of nodes per level. Taking into account the reduction possibilities in Q 07: Fourier decision diagrams may offer more eficient representations that MTDDs. The following example illustrates the use of Fourier decision trees on both Abelian and non-Abelian groups in representation of discrete functions. Example 4.3 A function f defined on a set of six points can be represented by the truth-vector F = [f(0),f (l), f (2), f (3), f (4), .f(5)]". In a direct graphical representation F is shown in Figure 4.15. Thisfigure corresponds to the truth-vector representation o f f over the cyclic group of order 6.
102
OPTIMIZATION OF DECISION DIAGRAMS
Thisfunction can be considered as a function f
( 2 1 ~ x 2on )
the group G6
= C2 x
@
Gs, where G3 = ({0,1,2},3). In that case it can be represented by the generalized multi-terminal decision tree shown in Figure 4.16. The group representations of Cz are the discrete Walshfunctions and, thus are given by the basic Walsh matrix
W(l)=A[ 2 1 -11 1 The group representations of G3 are given by the matrix
1 el
e2
where el = - !j (1 - i&) and e2 = - (1 + a&). Therefore, the Fourier transform on the Abelian group Gg = C2 x G3 is defined by the matrix
where 63 denotes the Kronecker product. In that case, f can be represented by the Fourier decision tree shown in Figure 4.17. This decision tree represents function
The function f can be considered as a function on the symmetric group of permutations Ss and in that case can be represented by the Fourier decision tree shown in Figure 4.18. This tree represents the function
f = ROSS (0) + RlSf (1)+ T 7 - ( R 2 S j ( 2 ) ) . The value of the constant node corresponding to S f ( 2 ) is a (2 x 2 ) matrix. Thus, this tree is a matrix-valued tree. If matrix-valued constant nodes are not allowed, S f ( 2 ) may be also represented by a tree. For that reason, we write S f ( 2 ) as a function q on the cyclic group G4 and represent it by the truth-vector Q = [S, ( 0 , O ) ,S j , Sg‘o’,S y ’ l ) ] TIn . that way, S2 can be represented by the MTBDDs on G4. The corresponding Fourier decision tree is a complex-valued decision tree. This tree is shown in Figure 4.19. From thisfigure, even with complex-valued Fourier decision tree, we still have savings in non-terminal nodes comparing to the decision trees based on Abelian groups (Figure 4.16 and 4.17). We also have further possibilities for the reduction of this tree. Depending on the values of elements of S j ( 2 ) we can use the Fourier decision tree on Gq to represent it and to save eventually some of the nodes representing elements of S j ( 2 ) . The corresponding decision trees representation o f f is as in Figure 4.20. The values of constant nodes are the Fourier coeficients S, of q on G4.
FOURIER DECISION DIAGRAMS
Fig. 4.15 One-level multiterminal decision tree of f in Example 4.3 on G G .
Fig. 4.76 Two-level multiterminal decision tree of f in Example 4.3 on GG= CZx Gs.
1-2\
FA
Fig. 4.1 7 Two-level Fourier decision tree of f in Example 4.3 on the Abelian group G6 = Cz x G3.
Fig. 4-78 One-level Fourier decision tree of f in Example4.3 on non-Abelian group S 3 .
103
104
OPTIMIZATION OF DECISION DIAGRAMS
fig. 4.19 Complex-valued Fourier DT on S3 with MTBDD for S f ( 2 ) .
Fig. 4.20 Complex-valued Fourier DT on S3 with Fourier DT for S f ( 2 ) .
The following example further elaborates the use of non-Abelian groups in optimization of decision diagrams.
Example4.4 A function f dejned on a set of 36 points can be considered as a function on a group G36 of order g = 36. This group can be decomposed in several different wajs, each providing a decision tree on Gs6 with different number of levels and nodes per levels. Figure 4.21 shows the decision tree on G36 corresponding to the decomposition G36 = Cz x C2 x C, x Cs, where C, denotes a cyclic group of order i. The DT consists of 19 non-terminal nodes. Figure 4.22 shows the decision tree corresponding to the same decomposition, but with different order ofthe constituent subgroups G36 = C3 x C:<x Cz x C2. This DT consists of 31 non-terminal nodes. Reduction possibilities in decision trees in Figure 4.21 and Figure 4.22 depend on thc subvectors of orders 2k and 3k for nodes with two and three outgoing edges, respectively. Fourier decision trees on finite non-Abelian groups (FNADDs) take advantage of the matrix notation of group representations and, therefore, the spectral coefficients. The cardinality of the dual object I' of a non-Abelian group G is always smaller than the order y of G. The same is true if the cardinalities of the dual objects F a and the orders of the constituent subgroups G, are compared. Note that in the matrix notation the group representations of G are generated through the generalized Kronecker product of the group representations of the constituent subgroups G,. Therefore, the
FOURIER DECISION DIAGRAMS
Fig. 4.21 Decision tree on G36
= CZ x
Fig. 4.22 Decision tree on G36 = Ga x
CZ x C3 x C3.
C3
x Cz x C2
105
106
OPTIMIZATION OF DECISION DIAGRAMS /
Fig. 4.23 Complex-valued FNADT on G36
= G2 x C3
x
S3.
number of nodes at intermediate levels is determined by the cardinality K , of I?,. It follows that some non-terminal nodes may be saved. In that way, FNADDs permit the reduction of both the number of levels and the number of nodes per levels in a decision tree. Unlike decision trees on Abelian groups, in FNADDs the use of large subgroups to reduce the number of levels does not increase the number of nodes per levels, since always K , < 9%.With that motivation, Fourier decision diagrams on finite non-Abelian groups were introduced in [22]. The following example illustrates the reduction of the number of non-terminal nodes per levels by using FNADDs.
Example 4.5 Consider FNADTs f o r the decomposition of G36 by using non-Abelian groups. In Figure 4.23, G36 = C, x C, x S3, where Ss denotes the symmetric group of permutation of order 3 defined in Example 2.2. The FNADT corresponding to the decomposition G36 = S3 x Ss is shown in Figure 4.24. In this figure, the MTDDs on G4 are used to represent the matrix-valued nodes of order 2 and the complex-FNADT on GIG= C, x Q2 f o r the node S f ( 8 ) with MTDTs on G4 to represent the matrix-valued nodes in this sub-tree. This DT consists of 13 non-terminal nodes and the disadvantage is the incremented number of levels to 6 instead of 5. Compared to decision trees in Figure 4.21 and Figure 4.22, reduction possibilities in Figure 4.23 and Figure 4.24 are diferent. Example 4.5 shows that by using decision trees on non-Abelian groups, the number of non-terminal nodes may be drastically reduced without essentially decreasing reduction possibilities in the decision tree.
FOURIER DECISION DIAGRAMS
107
Fig. 4.24 Complex-valued FNADT on G36 = S3 x S3.
4.3.2
Fourier decision diagrams
Fourier decision diagrams are derived by reducing Fourier decision trees. The reduction is done by using the generalized BDD reduction rules introduced for the reduction of WDDs [27].
Definition 4.3 (Generalized BDD reduction rules)
1. Delete all the redundant nodes where both edges point to the same node and connect the incoming edges of the deleted nodes to the corresponding successors. Relabel these edges as is shown in Figure 4.25(a). 2. Share all the equivalent sub-graphs, Figure 4.25( b). Definition 4.4 Fourier decision diagrams are decision diagrams derived from the Fourier decision trees by using the generalized BDD reduction rules. A Fourier decision diagram is reduced iffurther reduction with the same rules is impossible. The discussion in Section 4.1 about the reduction possibilities applies to Fourier decision trees. The main difference with MTDDs is that the reduction possibilities in the Fourier decision tree depend on the relationships among the values of Fourier coefficients Sf
108
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.25 Generalized BDD reduction rules.
of f . The order of the reduction possibility is determined by the cardinality K, of I', instead of gz of G , . Thus, in Fourier decision diagrams on non-Abelian groups (FNADDs) we have smaller reduction possibilities. If for a given assignment of values of a variable 2 , = p z ,p , E (0, . . . ,g,}, there are equal subvectors of orders gk,k < i, in the vector [Sf]of the Fourier coefficients of f , then the corresponding nodes at the k-th level may be joined. Thus, the redundant nodes may be deleted. If the equal subvectors of orders g k , k < i, in [Sf]correspond to different assignments of values of x,,the corresponding nodes may be shared.
4.4
DISCUSSION OF DIFFERENT DECOMPOSITIONS
To estimate the complexity of a decision diagram representation for a given function f , the size of a decision diagram is defined as the number of nodes in the decision diagram. In decision diagrams based calculation procedures, some calculations should be performed at each non-terminal node. The time complexity of the calculation procedure is, therefore, expressed as the size of decision diagram times the processing time per node. In some procedures where the processing of nodes differ, for example for nodes at different levels, we use the maximal or possibly average time per node. However, two additional parameters are important if wc consider the complexity of functions realizations derived from their decision diagram representations: 1. area that the produced network occupies,
2. propagation delay. Therefore, we should speak about the space-time complexity. In decision diagram based realbations, the time complcxity depends on the number of levels in the decision diagram. To estimate the space complexity, we can use the product of the number of levels and the number of nodes per level (the maximal number for the worst case and the average number of nodes per level otherwise). Therefore, the relevant parameters to estimate the spacc-time complexity of decision diagram based realizations for a given function f are
DISCUSSION OF DIFFERENT DECOMPOSITIONS
109
Table 4.2 Space-time complexity of DTs on GIG.
Decomposition
c, x c, x c2 x c,
Space-time complexity
2.2.2.2
c 4
x
c 2
x C8
4.4 2.8
x Q2
2.2.4 2 . (4 + 1 . 4 )
c 4
c, x c, x c 4 Q2
=2
. 4+2 . 1 . 4
1. number of levels, 2. maximal number of nodes per level, 3. average number of nodes per level. Example 4.9 compares these parameters for the 3-bit multiplier representation by MTDDs and Fourier decision diagrams for different decompositions of the domain group. However, to study the effect of different decoding of variables, or in the grouptheoretic interpretation, the effect of different decompositions of G, we normalized the space-time complexity with respect to the order of G. In this way, the decision trees in Figure 4.21, 4.22, 4.23 and 4.24 have all the normalized space-time complexity equal 16, since they all represent functions whose truth-vectors are of order 16. We represent the time complexity (expressed through the number of levels) by the number of factors in the decomposition of 16. We represent the space complexity of a level by the number of outgoing edges of the nodes the level consists of. Table 4.2 shows the space-time complexity of decision trees for different decompositions of GIG.In the first three decompositions we can speak about the multiplicative time complexities of orders 4, 2 and 2, respectively (number of factors). The space complexities are 2, 4 and 8. With the Fourier decision diagram we have the decomposition of the complexity into two additive parts. The time complexities of these parts arc 2 and 3, respectively. The space complexities of them are 2, 5 and 4, that is simpler than in the other cases where the time complexity is equal to 2. The decomposition of the space-time complexity in the Fourier decision diagram into two additive parts explains that we can save a non-terminal node in the replacement of the subtree of order 8 in Figure 4.6(a) and Figure 4.8 by the Fourier decision diagram on Q 2 . Combination of Abelian and non-Abelian groups of different orders provides decision diagrams of different complexity with respect to both number of levels and nodes. A suitable combination can provide an optimal representation with respect to both criteria. FNADDs permit a two-step optimization. First, we can use large non-Abelian subgroups G, to define a decision tree with a required number of levels and to reduce the number of non-terminal nodes per levels at the same time. Then, we can represent
110
OPTIMIZATION OF DECISION DIAGRAMS
the matrix-valued constant nodes by decision trees on Abelian groups and do the optimization by choosing among various possible decision trees on different Abelian groups.
4.4.1 Algorithm for optimization of DDs From the above consideration the following algorithm for the optimization of decision diagrams representations of discrete functions may be formulated.
Algorithm f o r optimization of FNADDs representations 1. Given a function f defined on a set D of d elements. Assume f o r D the structure of a group G of order g = d. 2. Determine the required number n of levels in the decision diagrams representation o f f . Determine possible decompositions of G into n subgroups G,. 3. Try different group structures f o r each G, by using the non-Abelian groups f o r the largest G, 's. Druw the matrix-valued FNADDsfor each possible combination of subgroups and calculate the number of non-terminal nodes and constant nodes. Chose the smallest size DD with the most planar structure. 4. End of the algorithm, if matrix-valued FNADDs are allowed. Otherwise, try representations of matrix-valued constant nodes by corresponding generalizations of MTBDDs and FNADDs. Chose the optimal representation f o r each node. To reduce the number of possible combinations of various decompositions of G and to get a quick solution, it is a reasonable to use decomposition into the subgroups of increasing orders. This view is based upon the previous experience in the closely related problem of efficient calculation of the Fourier transform on finite non-Abelian groups PI, [ l a .
4.5
REPRESENTATION OF TWO-VARIABLE FUNCTION GENERATOR
Therepresentationofthefunctionfgivenby thelruth-vectorF = [0,1,2,3,4,5,6,7IT was used in [ 131 as an example to illustrate the efficiency of Edge-valued Binary Decision diagrams (EVBDDs). We will use quite a similar example to illustrate the efficiency of FNADDs. Notice that EVBDDs are decision diagrams with attributed edges [25], which represent functions in the form of the arithmetic polynomials 1211. Arithmetic spectral transform decision diagrams (ACDDs) [27] do not have attributed edges, the same as FNADDs, but also represent functions in the form of arithmetic polynomials. Therefore, we will provide comparisons with ACDDs. Notice that definitions of the Arithmetic transform and arithmetic polynomials are considered in Section 5.3, and
REPRESENTATION OF lWO-VARIABLE FUNCTION GENERATOR
11 1
f
Fig. 4.26 Arithmetic transform decision diagram for f in the Example 4.6.
here we briefly recall the definition of the Arithmetic spectral transform decision diagrams.
Definition 4.5 1271 Arithmetic spectral transform decision diagrams (ACDDs) are binary decision diagrams, i.e., consists of nodes with two outgoing edges, where the values of constant nodes are the arithmetic spectral coeflcients. and labels at the edges are 1 and 2,. Example 4.6 Fora function f given by the vector F = [0,1,2,4,5,6,7IT,the arithmetic spectrum is calculated by using the arithmetic transform matrix for n = 3 defined in Example 5.3, thus, it is Af
= =
A(3)F [0,1,2,0,4,0,0,01T.
Figure 4.26 shows the ACDD for f . Example 4.7 Realization of switching functions by logic networks containing generators of all switching functions as basic design modules has been discussed already by Shannon 1201. A two-variablefunctions generator is eficiently applied as a design module also in present technologies, see for example [14], 1151. Figure 4.27 and Figure 4.28 show two-variable function generators in the notation by Shannon and with more recent symbols for elementary logic circuits [15]. The outputs of the two-variablefunction generator is described by the set of switching functions {xy, xy, x,Zy, y, x @ y, x V y, % . g, 2 @ y, G, x V Q, %, Z V y, It: V Q}. We
112
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.27 Two-variable function generator in Shannon notations. A .B A .B
A A.B
B AOB AvB
A .B
AOB -
B
A -vB
A
Av B -.
-
AvB
Fig. 4.28 Two-variable function generator.
REPRESENTATION OF WO-VARIABLE FUNCTION GENERATOR
~
1
1~2~
41
( 0
81
Fig. 4.29 ACDD representation of f ( z ) for TVFG.
If’ W, 1
,
1-2x,
Fig. 4.30 Matrix-valuedFNADD representation of
f(2)
for TVFG.
113
114
OPTIMIZATION OF DECISION DIAGRAMS
represent the truth-vectors of these functions by their decimal equivalents. Together with the constants 0 and I , these functions can be represented as an integer function f given by the truth-vector
F
=
[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]'.
Therefore, a MTBDD representing this f is the complete Shannon tree for switching functions of four variables. Thus, this representation o f f requires 31 nodes in 5 levels with 15 non-terminal nodes. The arithmetic spectrum o f f is Af
=
P, 1 , 2 , 0 , 4 , 0 , 0 , 0 , 8 , 0 , 0 , 0 , 0 , 0 , 0OI', ,
and it follows that f can be represented by the Arithmetic spectral transform decision diagram (ACDD) shown in Figure 4.29. This representation requires 15 nodes in 5 levels with 10 non-terminal nodes. I f f is considered as a function on GIG= C2 x Q 2 its Fourier spectrum is given by S f = &[120,-8, -32,0, S f ( 4 ) ,-64,0,0,0, S f ( 9 ) ] ' , with
Therefore, f can be represented by the matrix-valued FNADD in Figure 4.30. The representation requires 9 nodes in 3 levels with 3 non-terminal nodes. The matrix-valued node Sf(4) can be represented by the MTBDD on G4 shown in Figure 4.31. The impact of the zero matrix for Sf ( 9 ) is equal to the ordinary zerovalued constant node. By using this MTBDD as a sub-tree, f can be represented by the complex-valued FNADD shown in Figure 4.32. This decision diagram requires 12 nodes in 4 levels with 4 non-terminal nodes. The complexities of these different representations of two-variable function generator are compared in Table 4.3 showing the number of levels, non-terminal nodes (ntn), constant nodes (en) and the size of the MTBDD defined as the number of non-terminal and constant nodes.
4.6
REPRESENTATION OF ADDERS BY FOURIER DD
In this section we illustrate the application of FNADDs to the representation of an n-bit adder. It is shown in [27] that the number of nodes to represent n-bit adder by a MTBDD is 0 ( 2 n ) and by an ACDD O ( n ) .The number of nodes to represent n-bit multiplier by a MTBDD is O(4") and by an ACDD O ( n 2 ) .The number of levels in MTBDDs and ACDDs is always equal to n. It is shown on the example of 2-bit adders and multipliers that FNADDs permit some considerable savings in both number of levels and non-terminal nodes. Besides
REPRESENTATION OF ADDERS BY FOURIER DD
115
Table 4.3 Complexity of representation of TVFG in terms of the number of levels, non-terminal nodes (ntn), constant nodes (cn) and sizes.
Number of MTBDD ACDD Matrix-valued FNADD Complex-valued WADD
levels
ntn
5
15
cn 16
5
10
3 4
size
5
31 15
Figure 4.29
3
6
9
Figure4.30
4
8
12
Figure4.32
Fig. 4.31 MTBDD representationof Sf(4) for TVFG.
Fig. 4.32 Complex-valued FNADD representationof f for TVFG.
7 76
OPTlMlZATlON OF DEClSlON DlAGRAMS
Fig. 4.33 MTBDD representationof f for 2-bit adder.
J' A
I
Fig. 4.34 ACDD representationof f for 2-bit adder.
REPRESENTATION OF MULTIPLIERS BY FOURIER DD
117
that, the savings in the total number of nodes compared to MTBDDs representations are achieved even with complex-valued FNADDs. In that case the total number of nodes is equal to that used with ACDDs.
Example 4.8 The outputs of the 2-bit adder may be represented by the integer function f given by the truth-vector F = [ 0 , 2 , 2 , 4 , 1 , 3 , 3 , 5 , 1 , 3 , 3 , 5 , 2 , 4 , 46IT , [14]. Therefore, it may be represented by the MTBDD shown in Figure 4.33. The representation requires 19 nodes over 5 levels. The arithmetic spectrum o f f is Af = [0,2,2,0,1,0,0,0,1,0,0,0,0,0,0,0]'. Therefore, f may be represented by ACDD shown in Figure 4.34. The representation requires I I nodes in 5 levels. I f f is considered as a function on theJinite non-Abelian group GIG= Cz x Q2, where C;Lis the basic dyadic group and Q2 is the quaternion group, the Fourier spectrum o f f is
where Sf(4) =
-l+i 1-2
-1-2 -1-i
1
and Sf(9) =
Therefore, it may be represented by the matrix-valued Fourier decision tree shown in Figure 4.35, that can be reduced into the matrix-valued Fourier decision diagrams shown in Figure 4.36. The representation requires 8 nodes in 3 levels. The value of the constant nodes Sf (4) and Sf(9)are ( 2 x 2 ) matrices. The impact of the zero matrix for Sf(9)is equal to the ordinary zero-valued constant node. I f matrix-valued constant nodes are not allowed, that matrix can be considered as thefunction S f = [-1+ i, 1- i , -1 - i , -1 - z ] ~ Ifthis . function S f is consideredas a function on the cyclic group of order 4, Gq, it can be represented by the quaternary decision diagram QDD [18/ shown in Figure 4.37. By using this decision diagram, the complex-valued FNADD representing f jbr 2-bit adder may be derived as in Figure 4.38. The complexities of representing the 2-bit adder with these various decision diagrams are compared in Table 4.4, where number of levels, non-terminal nodes (ntn), constant nodes (en), and sizes of decision diagrams are shown.
4.7
REPRESENTATION OF MULTIPLIERS BY FOURIER DD
The above method is applied to the optimization of the representation of multipliers by Fourier decision diagrams. For simplicity, the presentation is given on the examples of 2-bit and %bit multipliers. Similar conclusions may be derived in the general case. The following example considers the efficiency of representing the 2-bit multiplier by different decision diagrams.
118
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.35 Fourier DT of f for 2-bit adder,
.f W 1
-
,
1-2x,
,
Fig. 4.36 Matrix-valued FNADD of f for 2-bit adder.
REPRESENTATION OF MULTIPLIERS BY FOURIER DD
/
',
I '
!-'+l
Fig. 4.37 MTBDD representationof Sf(4) for 2-bit adder.
1-2 x,
,
~-
fig. 4.38 Complex-valued FNADD of f for 2-bit adder.
119
720
OPTIMIZATION OF DECISION DIAGRAMS
Table 4.4 Complexities of representing the 2-bit adder in terms of levels, number of non-terminal nodes (ntn), constant nodes (cn), and sizes.
Number of
ntn
levels
MTBDD ACDD
Figure 4.34
Matrix-valued FNADD
Figure 4.36
~
Complex- val ued FNADD
6
10
Figure4.38
Example 4.9 The outputs of the 2-bit multiplier may be represented by the integer function f given by the truth-vector (141
F
=
[O,O,0,4,0,0,2,6,0,2,0,6,1,3,3,9IT.
The MTBDD representation off requires 20 nodes over 5 levels. The arithmetic spectrum o f f is Af = [O, O , O , 4,0,0,2,0,0,2,0,0,1,0,0,0]'. Therefore f may be represented by an arithmetic spectral transform decision diagrams (ACDD) with 1.5 nodes in 5' levels. Edge-valued binary decision diagrams (EVBDD) [131 representation of f for the 2-bit multiplier requires I0 non-terminal nodes over 5 levels with I 1 weighting coejticients. I f f is considered as a function on the finite non-Abelian group G16 = C, x Q2, where C, is the basic dyadic group and Q 2 is the quaternion group, the Fourier spectrum o f f ( z ) is
1
1
162 -8-242 and Sf (9) = 8 - 24i -162 J L J Therefore, it may be represented by the matrix-valued Fourier decision diugruni which requires 8 nodes in 3 levels. The values of the constant nodes Sf(4) and Sf(9) are ( 2 x 2 ) matrices. Ifthe matrix-valued constant nodes are not allowed, that matrix can be considered asthefunction Sf = [16i,8-242, - 8 - 2 4 , -1GiIT. Ifthisfunction S f isconsidered as a function on the cyclic group of order 4, G4, it can be represented by a quaternary decision diagram QDD [18]. By using this decision diagram, the complex-valued FNADD may be derived. This decision diagram representation of the 2-bit niultiplier requires 13 nodes in 4 levels. The complexities of representing the 2-bit multiplier with these various decision diagrams are compared in Table 4.5.
where Sf(4)
=
L
REPRESENTATION OF MULTIPLIERS BY FOURIER DD
121
Table 4.5 Complexities of representing the 2-bit multiplier in terms of the number of levels, non-terminal nodes (ntn), constant nodes (cn) and sizes.
Number of MTBDD
levels
ntn
cn
size
5
13
7
20
ACDD
I
5
I11141151
EVBDD
I
5
I101111211
FNADD Complex-valued FNADD
Fig. 4.39 MTDD for the 3-bit multiplier on CG4= C4 x C4 x C4.
In the above example, the reduction of levels is achieved by using Fourier decision diagrams. The other decision diagrams on Abelian groups GIG= C, x C, x CZx C, do not offer such possibility. The reduction of levels may be achieved with QDDs and their corresponding generalizations for different decompositions of the domain group G for f . The following example studies the effect of different decompositions of the domain group G64 in the representation of 3-bit multiplier with MTDDs and compares these representations with a Fourier decision diagram representation.
122
OPTIMIZATION OF DECISION DIAGRAMS
Example 4.10 The outputs of the 3-bit multiplier may be represented by the integer function f given by the truth-vector [14] F
=
[0,0,0,0,0,0,0,0,0,1,2,3,4,5,6,7, 0,2,4,6,8,10,12,14,0,3,6,9,12,15,18,21, 0,4,8,12,16,20,24,28,0,5,10,15,20,25,30,35, 0,6,12,18,24,30,36,42,0,7,14,21,28,35,42,49IT.
I f f is considered as a function on thefinite non-Abelian group G64 = Q2 x where &a is the quaternion group, the Fourier spectrum off is Sf
1
Q2,
1 -[49, -7,28,0,Sf(4), -7,1,4,0,Sf(9) 4 -28,4,16,0,Sj(14),0,0,0,0, Sf(19)
s f ( 2 0 ) >Sf (211, Sf (22), Sf(23) sf(24)lr. I
Sf(9)
Sf(21) =
1
Sf(14) = Sf(22) = 4
Sf(24) = 2
I
-l+i i
1-2
-l+i 1-2 -l+i
-l+i l+i
I
'
-l+i l+i
1
'
1 l+i 1-2 1 -1-i -1 l + i
-l+i We represent the matrix-valued nodes simply by MTDTs of the corresponding orders by taking the advantages from the properties of the matrix-valued Fourier coefficients'. In that way we produce the complex-valued Fourier decision diagrams representing the 3-bit multiplier [24/, [25]. The complexities of representing the -?-bit multiplier with MTDDs for different decompositions of G64 and complex-valued Fourier DD on G64 = Qz x Q2 are compared in Table 4.6. This table shows the number of levels, non-terminal nodes (ntn),muximum number of nodes per level (max),minimum number of nodes per level (min), average number of nodes per level (av), number of constant nodes(cn), total of nodes, i.e., the size (s), and the maximum number of edges per node (e). Fourier decision diagrams permit the use of nodes with negated edges (FNADD n.e. in Table 4.6) [24/, [25], which means that the number of constant nodes may be
' In many cases, matrix-valued Fourier coefficients are symmetric, Hermitean or skew-Hermitean matrices.
COMPLEXITYOF FNADD
123
considerably reduced. In this example, however; the negative edges can not be used with MTDDs. Figure 4.39 shows an MTDD for the 3-bit multiplierfor the decomposition G64 = C, x C, x C,. Figure 4.40 shows an FNADD for the 3-bit multiplierfor the decomposition Gs4 = Q 2 x QZ [241, [251.
Table 4.6 Complexities of representing the 3-bit multiplier in terms of the number of levels, non-terminal nodes (ntn), maximum (max) and minimum (min) number of nodes per level, average number per level (av), constant nodes (cn), sizes (s), and maximum number of edges per level (e).
4.8
COMPLEXITY OF FNADD
In this section, we compare various decision diagrams on dyadic groups with FNADDs on the quaternion groups. Multiple-output functions are represented by Shared BDDs (SBDDs), Multi-terminal binary decision diagrams (MTBDDs) [ 141, and FNADDs. For representation by MTBDDs and FNADDs, the benchmark functions are first represented by the integer-valued functions f . Depending on the value of R, in FNADDs on quaternion groups CzQ; or C,Q;, we use as the root node the Shannon nodes with two or four outgoing edges. Table 4.7 compares sizes (s) and widths (w) of SBDDs and FNADDs for various benchmark functions. For FNADDs the values of non-terminal nodes (ntn) and constant nodes (cn) are shown separately. Thus, the size of FNADDs is the sum of these
124
OPTIMIZATION OF DECISION DIAGRAMS
S,=1 i
S,=l+i
-3
0 1
6
h=w3 +w3
Fig. 4.40
2
13
3
4
15
c=w3 +w3 +w3
d v 3 fw3
5
8
10
e=w3 +w3 +w3
FNADD for the 3-bit multiplier on Gfi4 = Q2 x
9
11
p3 +w3 12
14
h=w3 +w3
Q2.
TTTT
Table 4.7 SBDDs and FNADDs for benchmark functions.
-
.f 5xpl
SBDD S
90
con 1
I16 20
rd53
25
rd73 xor5
45 11 -
bw
a
ntn
cn
s
w
a
25
2250
39
128
167
18
3006
37
4292
9
0.36
25
7
14
21
450 22
23
30
53
5
6
I1
4 6 3 6 2
136
13
25 12
34
1 00 150
150 63 318 22
1.83 0.50 0.70 0.83
W -
5 6 10 2 -
rn -
0.35
two values. It is also shown the ratio of non-terminal and constat nodes T , = ntn/cn. This table shows that, besides depth reduction, we get width reduction, except for 5xpl. Except this function and conl, the area is also reduced. In FNADDs, the ratio of non-terminal and constant nodes T, is quite more convenient. For example, in xor5, where the size and width remain the same, this ratio T, is 5.5 and 0.83, respectively. Table 4.8 compares sizes of SBDDs and FNADDs for adders and multipliers. In this example, FNADDs provide the reduction of all three parameters, depth, size, and width and the reduction is considerable compared to SBDDs. The reduction possibilities in these decision diagrams are compared as the percentage of nodes used from the total of nodes in the corresponding decision trees. It may be seen that these values are comparable, thus, the possibility to do reduction in SBDDs and FNADDs is comparable. In FNADDs, the reduction of area is considerable and shows the advantage of them. Table 4.9 compares sizes of SBDDs, MTBDDs and FNADDs for adders and multipliers. In SBDDs and MTBDDs, the domain group is C;. The domain group for FNADDs is shown in the table, since depends on n. We can observe that, besides depth reduction, FNADDs permit reduction of width and size. At the same timc, FNADDs have smaller area and convenient ratio of non-terminal and constant nodes for applications where reduction of the number of non-terminal nodes at the price of constant nodes may be required. When n increases, these properties of FNADDs increase. Table 4.10 compares the FNADDs to some other decision diagrams based on spectral transforms. The Arithmetic transform DDs (ACDDs), Walsh transform DDs (WDDs) in (1,- 1)coding (271, and Complex-HadamardTransfom DDs [71 arc compared with FNADDs for some benchmark functions. ACDDs, WDDs, and CHTDDs arc DDs on Abelian groups. Therefore, the depth is equal to n. For rd53 and xor5, the sizes of FNADDs are equal to those of other DDs. However, the number of nonterminal nodes and the width are reduced. In other cases in this example, FNADDs are more efficient with respect to depth, width and ske. However, as for other decision diagrams, examples where FNADDs are not efficient certainly can be found. As an ex-
126
OPTIMIZATION OF DECISION DIAGRAMS
i
Table 4.8 SBDDs and FNADDs for adders and multipliers.
adders SBDD 12
s
w
%
a
ntn
2 3 4 5 6
8 57 103 226
21 20 30 62
22.00 42.86 19.84 10.98
168 1140 3090 14012
4
477
126
5.81
cn -
7 6 7 14 14 18 16 21 12 60 102 -
multipliers
I
FNADD -
S
W
-
11 2 52.38 4 3 1.70 13 7 33.73 28 7 8.31 34 7 4.00 33 FNADD
a 95 945 620 1 53912 151296
%
94
22
22
0.57
52 196 238 231
0.85 1.00 1.12 1.75
COMPLEXITY OF FNADD
Table 4.9 SBDDs, MTBDDs and FNADDs for adders and multipliers.
cn
I
adders s
W
20 14 4 30 30 7 62 62 7 127 12
832 33
126 I26 7
a 1140 784 52 3090 4320 203 14012 2 1824 238 60102 104832 23 1
27.5 2.73 0.86
50.50 3.64 0.93 112.00 4.59 1.12 228.50 5.55 I .75
j-z
multipliers
rn -
116
30.5 2.15 0.45
159
39 120 11
6201 39600 770
78.50 2.67 0.52
47 3
114 496 23
53922 660672 2392
235.50 2.92 0.79
192 2016 22
151296 10624320 2068
393.00 3.257 0.92
46
1238 4:
1
5270 :7:
127
128
OPTIMIZATION OF DECISION DIAGRAMS
Table 4.10 Sizes of ACDDs, WDDs, CHTDDs and FNADDs.
-
f 5xp 1 bw con I rd53 rd73 xor5
-
ACDD -- ntn cn S W -
CHTDDs -
L
38 31
11
32
40 5 15 6 27 6 6 15 -
L_
a
49 63
10
16
490 I008
45 21
10 5
450 1 05
33 21
6 5
188 105
--
ntn rn -
3.45 0.97
I27 31
8 2.5
I15 15
27 4.5 9 2.5 -
1,-1) __ - I - - WDD( cn ntn S rn f __ - 5xp I 37 13 50 2.85 31 32 63 16 bw 008 0.97 48 1 1 59 13 con 1 767 4.36 rd53 15 5 20 5 100 3.00 28 23 I 5 33 7 rd73 5.6 5 7 I 2 I xor5 7 2.50 -
Table 4.11
n=2r 4 6
8
cn
S
128 32 42 6
255 63
8 2
W -
a
157 21
64 16 56 5
16320 I008 8792 105
36 11
7 2
252 22
---
FNADD __ S
W
167 34 25 21
18 4 6 3 6 2
53 II
a 3006 136 150 63 318 22
0.99 0.96 2.74
2.5 3.37 4.5
-
rn __
0.30 0.36 1.08 0.50 0.77 0.83
BDDs and FNADDs for Achilles’ heel functions.
BDD worst best 8 12 14
6 8 10
worst
13 37 71
FNADD best decomposition 12
C2 Q2
25 40
C4Q;
Q$
ample consider the Achille’s heel function f = 21 y1 V x 2 y V ~ -. Table 4.1 I shows the sizes of FNADDs for Achille’s heel function for different values of n and for two different orderings X I ,22,. . . , xa,,y1, y2,. . . , ~2~ and XI,y1,x2, y 2 , . . . ,xnyTL. This example shows that, as in any other decision diagrams, the sizes of FNADDs greatly depends on the ordering of variables. In this example, the depth reduction in FNADDs is achieved at the price of the increase the size of the FNADDs.
FOURIER DDS WITH PREPROCESSING
4.9
129
FOURIER DDS WITH PREPROCESSING
Fourier decision diagrams on non-Abelian groups take advantages from matrix representations of the Fourier transform. In this section, we propose a method which permits to further exploit these properties in decision diagram representations of discrete functions. We first convert agiven number-valued function f into a matrix-valued function f m by splitting the vector F of function values for f into subvectors which are arranged as rows of some matrices representing elements of a matrix-valued vector F,, defining f m . We represent f m by a matrix-valued FNADD. Then, in the thus derived FNADD, we represent each matrix-valued constant node by a Shannon decision diagram. In this way, we obtain a number-valued decision diagram representing f in a very compact form. In what follows, we define the matrix-valued functions and the related Fourier transform. Then, we define the Fourier decision diagrams for matrix-valued functions. This transform is used to define the Fourier decision diagrams for representation of matrix-valued functions. We explain by examples how these decision diagrams can be used to represent number-valued functions. 4.9.1
Matrix-valued functions
A discrete function f t P ( G ) is conveniently represented as a vector of its values at all the points of G, thus, f is given as F = [f(0), . . . , f (y - 1)IT.
Definition 4.6 (Matrix-valuedfunctions) A function f defined on G taking values in a set of ( u x b) matrices hI,,b over afield P is called a matrix-valued function on G. The space of all matrix-valued function on G over P is denoted by P,,b(G). A matrix-valued function f E P,~b(G)is conveniently represented by a vector F = [f(O). . . . , f ( y - 1)IT,where f ( i ) are ( u x b) matrices over P. In this section mainly functions whose values are square matrices are considered, thus, a = h = r , where r,, is a given number [ 1 I]. This dimension will be given the value r W where , T , is the dimension of the w-th unitary irreducible representation of G over P [ I I ] . In this vector notation, a number-valued discrete function can be transferred into its matrix-valued equivalent by using the following obvious algorithms.
Algorithm 4.1 (Vector to matrix transformation (by columns)) 1. Given a vector F of order y = n r i . Split F into subvectors consisting of r," successive elenients f ( i )of F.
2. Write as columns of a matrix the first r, successive subvectors Continue with the following r , subvectors until n matrices of order (I-?" x I-?,,) are generated. 3. The vector [F,,] of order gn = n whose elements are the thus generated matrices f ( i ) represents the matrix-valued function f v L corresponding to f .
130
OPTIMIZATION OF DECISION DIAGRAMS
Algorithm 4.2 (Vector to matrix transformation (by rows)) Algorithm 4.1, but with subvectors in 2 written as rows of the generated matrices. Conversely, an ( r wx r,) matrix can be considered as a function on a group of order r i and, thus, represented by a vector generated by using the following algorithms.
Algorithm 4.3 (Matrix to vector transformation (by columns)) 1. Given an ( r , x r,) matrix M over P. 2. Generate a vector V of order r i by concatenating the columns of M.
Algorithm 4.4 (Matrix to vector transformation (by rows)) Algorithm 4.3, but concatenating the rows of M in 2. 4.9.2
Fourier transform for matrix-valued functions
It is assumed that the conditions for the existence of the Fourier transform on G over P a r e satisfied in P,,, , l u (G). Denote by R,(z) the value of the w-th unitary irreducible representation R, of G over P at the point z E G. Thus, R,(z) stands for an r , by r, matrix with elements R$.')(J.)E P , and a,.j E ( 0 , . . . , r , - 1). The set of all unitary irreducible representations of G forms the dual object r of G. In the case of Abelian groups, all the unitary irreducible representations are one-dimensional and, thus, reduce to group characters. In that case r expresses the structure of a multiplicative group isomorphic with G. If G is decomposable in the form (2.4), then the cardinality K of r can be written as K = K,, where K , is the cardinality of the dual object r, of G,. The orthogonality relation for the components of the matrix functions of the dual object r is
ny=l
where t u , u E r, 1 I i , k I r,, 1 I j , r 5 r,. The bar denotes complex conjugation and 6 is the Kronecker delta. The character xw of R,, is xw(z) = W R w ( x ) ) .
The orthogonality relations for characters are
FOURIER DDS WITH PREPROCESSING
737
where p , is the number of elements in the conjugate class of G that contains 2. Recall that the number of conjugate classes of a group G is equal to the number of elements in the set I?.
Definition 4.7 The direct arzd inverse Fourier transform off E Pk,,(G) are defined respectively by g-1
S,(w) = (Sy”)(w))= r,g-l
f(”’j)(u)R,(u-l),
(4.1)
u=o
K-l
C Tr(S~’”(w)R,(e)),
f ( s ) = (f(””)(s)) =
(4.2)
W=O
where K is the number of unitary irreducible representations R, of G over P. Notice that for a function f E Pk ,(G) with group representations of order r,, the Fourier coefficients are in A&,, ,mr,, . Here and in the sequel we shall assume, without explicitly saying so, that all arithmetical operations are carried out in the field P. In the matrix notation, the Fourier transform matrix on finite groups is given as follows [I 1I.
Definition 4.8 The Fourier transform on G is defined by the transform matrix [R]-’ inverse to the matrix of basis functions defined as group representations for G, i.e., [R]= Y [ R ~ , R ~ , . . . R K where -~], ...
RIP“) (0)
R:”-l O)(g - 1) . . .
(9 - 1)
...
RY’”(0)
Rrtt pl>o)(g- 1) ...
,
. . RrZt - l % r u
(9 - 1)
... R K - ~= 1)
(T,‘
’ ’ ‘
-1,r,, - 1 )
RK-~
(9 - 1)
where [R]E Mg,T: and Rk’”(.) denotes the ( i , J ) - t helement ofR,(.), 1 I p , q F r,. Thus, R E Idg,,.
132
OPTIMIZATION OF DECISION DIAGRAMS
-:I,
Example 4.11 The Fourier transform on 5'3 over G F (11)introduced in Example 2.2 is given by
" 21'1
[R]-'=-
1 1 -1 2B 2A 2C
-1 2D
2E
with notations as in Table 2.2 and Table 2.5, respectively. Notice that in the case of representations over GF(11),the normalization factor is 116 = 2. Example 4.12 The Fourier transform for functions on Qz over C is defined by group representations in Table 2.5 and it was introduced in Example 2.3. For convenience, it is repeated here, thus, 1 1 1 1 1 1 1 1 -1 1 -1 1 -1 1 1 1 1 1 -1 -1 -1 1 -1 1 -1 -1 1 -1 21 2iB -21 2iA 2E 2iD 2C -2iD
with notations as in Table 2.9. Example 4.13 Function xor5 is afive variable function defined as xor5 = zl@a28 x3 @ xq @ x5 and is given by the truth-vector
F = [0,l,l,0,l,0,0,l,l,0,O,l,0,l,l,0,l,0,0,l,0,l,l,O,0,l,l,0,l,0,0,l]T. By using the algorithm 4.2, xor5 transfers into the matrix-valuedfunction f m given by the vector 'Fm'
=
":;I,[ [
1 0 0 1]9[;
1 0 0 I]>[;
&[
:' I .[
0 1 1 01.
0 1 1 0 1 0]&l Ill7
The Fourier coeficients are calculated as
+ + + + +f(z,'J)(G)Rw(4) + f("j)(7)RU,(5).
2 S~(L= U )(Sf")(w))= -(f('>j)(0)RW,(O) f"'3)(1)Rw(3)f'"J'(2)R,,(2)
8 + f ( z J)(3)Ru,(l)f ( z 3 J ) ( 4 ) R w ( G f("J)(5)Ru,(7) )
The Fourier spectrum of
fm
is given by
FOURIER DDS WITH PREPROCESSING
133
where
Sf(1)= S f ( 2 ) = Sf(3) =
1 Sf(4) = -
I
4 -
l-il
-Ibi -1-2
-l+i
[
-l+i]
lSi l+i
1-i
[
[:
: I 7
[
1
l$-i -l+i l+i 1-i
-1-2 -1-2
1-21 -l+i ’
or in matrix notation
[Sfl =
s1 -1-i l+i
l+i
-l+i -l+Z] I-i
[
l+i l+i -1-i -1
-1+i 1-i
1
1-i] -1Si
Thefunction can be reconstructed by the application ofthe relation 4.2 asfollows:
134
OPTIMIZATION OF DECISION DIAGRAMS
We will illustrate this procedure by calculating the value of the function f (0) and f(7).
++[ I,++[ ++[
Tr(4.1)
+Tr(0
'
+
1) Tr(O.1)
[
+ T r ( 0 .1)
-1-i 1-2 1 0 -1-2 0 11) 4 + O + 0 0 - 4 = 0, Tr(4.1) T r ( 0 . 1 ) T r ( 0 1) T r ( 0 . 1)
+ +
l+i
+
-l+i 1-21
l+i 4 + 0 + 0 + 0 + 4 = 8,
+
+
'
[
1 0 0 11)
+
Tr(4.1) Tr(O.1) Tr(O.1) l + i -1+i 1 l+i 1-21 0 4 + 0 + 0 + 0 + 4 = 8, T r ( 4 1) Tr(O.1) T r ( O . 1 ) '
+
+ T r ( 0 .1)
[
0 11)
+
+ T r ( 0 .1)
+
+
4+0+0+0-4=0
++[ ++[ ++[ ++[ +
Tr(4.1) T r ( 0 . (-1)) T r ( 0 . (-1)) T r ( 0 . 1 ) -1-2 1-2 0 i -1-2 -l+i] i 01) 4 + 0 + O + 0 + 4 = 8, Tr(4. (-1)) T ' r ( 0 . (-1)) T r ( 0 1) T r ( 0 . 1) l + i -1+i 0 i l+i 1 - i ] [ 2 01) 4 + 0 + 0 + 0 - 4 = 0, Tr(4.1) T r ( 0 . (-1)) T r ( 0 .(-1)) T'r(0 1) l + i -l+i 0 i l+i 1-21 i 01) 4 + 0 + 0 + 0 - 4 = 0, T r ( 4 . I) T r ( 0 . (-1)) T r ( 0 . (-1)) T ' r ( 0 . 1) -1-2 1-2 0 i -1-i -l+i] i 01) 4 + 0 + 0 0 + 4 = 8.
+
+
+
+
[
+
+
'
+
+
[
+
[
The other function values are determined in the same way.
+
'
FOURIER DECISION TREES WITH PREPROCESSING
4.1 0
135
FOURIER DECISION TREES WITH PREPROCESSING
Decision trees are defined by using some function expansions [ 161. For example, the Shannon tree is defined by using the Shannon decomposition rule [3]. Walsh transform decision trees (WDTs) are defined by using the function expansion derived from the Walsh transform [27]. The discrete Walsh transform is the Fourier transform on the dyadic groups, since the Walsh functions are the group characters of the dyadic groups. Thus, WDDs are an example of Fourier DDs on these particular Abelian groups. The concept was extended to any finite not necessarily Abelian group through Fourier DDs [22] that were obtained by using the function decomposition determined by the Fourier transform on finite groups to define the Fourier transform decision trees. In this section we extend the approach to matrix-valued functions. The Fourier transform on a finite group G decomposable in the form (2.4) may be considered as the n-dimensional Fourier transform on the constituent subgroups G,. The Fourier transforms on G, will be used to define a mapping performed at the nodes of a Fourier decision tree to associate a function to the DT.
Definition 4.9 Denote by S f the Fourier transform coefficients on afinite not necessarily Abelian group G representable in the form (2.4). The Fourier decision tree on G over P is defined as the decision tree in which the nodes at the i-th level represent functions f(x,)
=
(f‘””’(x,))=
C ?‘r(S~’”)(w,)R,~(xi)),xi W,
E
Gi.
Er,
This definition differs from Definition 4.1 in the property that all the Fourier coefficients are matrix-valued. With this definition, Fourier decision trees are the decision trees on G in which each path from the root node up to the constant nodes corresponds to a unitary irreducible representation of G. A node at the i-th level has Ki outgoing edges denoted by R w r ( j ) ,Ej ri. In the figures of decision trees we use the short notation w! for Rwt( j ) or simply j when the context does not allow misunderstandings. For a given f , the values of constant nodes in the Fourier decision diagrams are the Fourier coefficients o f f . It follows that, as in decision diagrams on Abelian groups, we perform the inverse Fourier transform in the determination o f f represented by a given Fourier decision tree by following the labels at the edges in the DT [211. In application of the thus defined Fourier decision trees to number-valued functions, these functions should be first transferred into their matrix-valued equivalents. Therefore, such decision trees will be called the Fourier decision trees on non-Abelian groups with preprocessing (FNAPDTs). The structure of matrix-valued FNAPDTs (mvFNAPDTs) is determined as follows. This statement points out the difference with decision trees on Abelian groups.
Observation 4.6 The number of levels in a FNAPDTs is equal to the number of constituent subgroups G iin the domain group G o f f . The number of outgoing edges of nodes at the i-th level is determined by the cardinality K , of the dual object rzof
136
OPTIMIZATION OF DECISION DIAGRAMS
the corresponding subgroup G,. The number of constant nodes in FNAPDTs is equal to the cardindig of the dual object r of G. The difference with the FNADTs is that in FNAPDTs all the constant nodes are matrix-valued. In FNADTs most of the constant nodes are number-valued. Those corresponding to group representations of orders T, > 1 are the matrix-valued. That difference originates form the difference in definition of the Fouricr transform for number-valued and matrix-valued functions on non-Abelian groups. The matrix-valued coefficients in FNADTs and FNAPDTs may be also represented by decision trees by using the method of representation of matrices by the decision diagrams [ 5 ] . In that way, we derive the integer-valued or complex-valued Fourier decision trees on finite non-Abelian groups, depending on the field P over which the group representations are taken. In that case, the number of constant nodes in a Fourier DT on a non-Abelian group is equal to g,, that is, it is the same as in any DT on Abelian group of the same order. Comparing to the matrix-valued Fourier decision trees, the number of levels in an integer-valued or complex-valued Fourier DT is increased to represent the matrix-valued constant nodes. However, comparing to decision trees on Abelian groups, non-terminal nodes still may be saved, since always K , < g 2 . The non-terminal nodes in the Fourier decision trees are denoted by F, for Abelian groups and on non-Abelian groups by FA, and FNA, for decision trees with preprocessing. The index i denotes the number of outgoing edges and is equal to the cardinality K , of the dual object of G,. 4.1 1 FOURIER DECISION DIAGRAMS WITH PREPROCESSING Decision diagrams are defined by the reduction of the decision trees. The reduction rules are chosen depending on the range of the functions represented. Therefore, the Fourier decision diagrums with preprocessing (FNAPDDs) are defined as follows. Definition 4.10 Fourier decision diagrams with preprocessing are the decision diagrams that are derived from the Fourier decision trees with preprocessing by using the generalized BDD reduction rules. A FNAPDDs is called reduced iffurther reduction with the same rules is impossible. By definition, Fourier decision trees are graphical representations of the Fourier expansion of a given function f . Therefore, they are canonic representations since the Fourier expansions are canonic representations. The same applics to Fourier decision diagrams, since the generalized BDD reduction rules do not destroy or diminish the information content in the decision trees. In decision diagram representations of integer-valued functions nodes with negative edges are usually used to reduce the complexity of a decision diagram. A negative edge indicatcs that the value or subfunction 'u represented by the node where the edge points to, should be multiplied by -1. Thus, two nodes with values L' and -0 and
CONSTRUCTION OF FNAPDD
137
Table 4.12 Labels of edges. V
I
1
y+iz
y - iz -y - iz
-y
+ iz
the incoming edges j and k , may be replaced by one node whose value is u and two incoming edges are labeled by j and -k. In various decision diagrams representing complex-valued functions on finite groups, as well as in Fourier decision diagrams over C for whatever integer or complex-valued function, there are nodes whose outgoing edges point to nodes representing pure imaginary values or nodes representing mutually complex-conjugate values. In these decision diagrams, by the same justification as for the nodes with negative edges, we can consider imaginary edges denoting the multiplication with the imaginary unity i. Thus, two nodes whose values are w and iv with the incoming edges labeled by j and k, respectively, can be represented by one node whose value is ti and the two incoming edges are labeled by j and ik. In a similar way, simplification of decision diagrams representations may be done by introduction of the node with complex-conjugate edges permitting to reduce nodes representing a value or subfunction v and its complex-conjugate I)*. Labels of corresponding imaginary and complex-conjugate edges are shown in Table 4.12. In this table, u denotes the value or subfunction represented by a node and 1 is the label at the edge corresponding to Rw,( j ) for non-Abelian and 2 , = j for Abelian groups. The last row in the table shows, that if there is a node showing the value z and another showing the imaginary value iz,which is pointed by an edge j , we can redirect j to the node z and change the label into i . j , where i =
n.
4.12
CONSTRUCTION OF FNAPDD
FNADDs take advantage from the matrix notation and thus provide compact representation off in the spectral domain. In FNAPDDs the same advantages are extended to both original and spectral domain. If a given function f on a group G of ordcr ,y is transferred into a matrix-valued function on a finite non-Abelian group G, of order glL = y/r,,, where T, is the greatest order of the group representations of G,,, then, f is represented by the FNAPDD on G,. The elements of the matrix-valued constant nodes in this FNAPDDs may be considered as functions on an Abelian group of order r , and represented by the
138
OPTIMIZATION OF DECISION DIAGRAMS
I'
0
Fig. 4.42 FNADD for xor5.
Fig. 4.41 BDD representation for xor5.
decision diagrams on this group. In that way a decision diagram with small depth and width may be produced for f . Construction of FNAPDDs is explained and these decision diagrams compared with FNADDs by the following examples. Example 4.14 Function xor5 is given by the truth-vector F
=
[O,l,l,O,l,O,O,l,l,O,O,l,O,l,l,O,l,O,O,l,O,l,l,O,O,l,l,~,l,O,O,l]T,
1; -;
and can be represented by the BDD in Figure 4.41. xor5 can be considered as a function on the group G32 = C4 x Q 2 . The group representations of C, over the complex field are given by the complex Walsh matrix 1 -11
CW =
Since, CW-' matrix
=
-1
I]
1 -1 -1 -2
iCW, the Fourier transform on G32 RYi
=
CW-'@Q:'.
= C4 x Q2
is defined by the
CONSTRUCTIONOF FNAPDD
.
Fig. 4.43
Fig. 4.44
Matrix-valued FNAPDD for xor5.
FNAPDD for XOr5 with elements of mv nodes.
139
140
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.45 Complex-valued FNAPDD for xor5.
The Fourier spectrum of xorS on this group is given by
where
Therefore, this function can be represented by the FNADD shown in Figure 4.42. In this figure, the number 118 before the rectangle around constant nodes denotes that all the values of constant nodes has to be scaled by 118. The same notational convention will be used in further examples. The Fourier spectrumf o r xorS viewed as a matrix-valuedfunction on Q 2 is determined in Example 4.13. I t follows that xorS may be represented by the matrix-valued FNAPDD in Figure 4.43. By using the Algorithm 4.2, the matrix-valued nodes in Figure 4.43 can be represented by Multi-terminal decision diagrams (MTDDs) on C: or Ci. From (4.2), to read the function values f o r f , we should work with matrix elements of the niatrixvalued Fourier spectral coeficients, i.e., with submatrices S;;") ('uI) of order r , . It follows, that in a mvFNAPDD, it is suflcient to represent the matrix-valued coefl-
CONSTRUCTION OF FNAPDD
141
Table 4.73 Complexity of representation of xor5 in terms of the number of levels, non-terminal nodes (ntn), constant nodes (cn), and sizes.
cients until their submatrix elements of order r,. Howevel; these elements can be further represented by complex-valued decision diagrams. Therefore, xor5 may be represented by the complex-valued FNAPDD in Figure 4.45. Complexity of representation of xor5 by BDD FNADD and FNAPDD is compured in Table 4.13. This table shows the depth (d), width (w), number of non-terminal nodes (ntn), constant nodes (cn), and the size ( s )for different decision diagrams for xor5.
Example 4.14 shows that compared to BDD and FNADD for xorfi, FNAPDD reduces the depth, width and size of the DD. The number of non-terminal nodes is also reduced. The price is the increased number of constant nodes. However, that saving of non-terminal nodes (7 instead 9 for BDD) is in some applications more important than the increase of the number of constant nodes (4 instead of 2). The following example illustrates the use of FNAPDDs when the domain group G is represented as the product of Abelian and non-Abelian groups.
Example 4.15 Figure4.46showsBDDfortheswitchingfunction .f(xl,x2,~3,54, ~ 5 given by the following truth-table (tukenfrom [14]).
)
142
OPTIMIZATION OF DECISION DIAGRAMS
Fig. 4.46 BDD for f in Example 4.15.
Fig. 4.47 Matrix-valued FNADD for f in Example 4.15.
CONSTRUCTION OF FNAPDD
I
f
’ FA,
Fig. 4.48 Complex-valued FNADD for f in Example 4.15.
Fig. 4.49 Matrix-valued FNAPDD for f in Example 4.15.
143
144
OPTIMIZATION OF DECISION DIAGRAMS
1.
f
Fig. 4.50 FNAPDD for f in Example 4.15 with elements of mv nodes.
Fig. 4.51
Complex-valued FNAPDD for f in Example 4.15.
CONSTRUCTION OF FNAPDD
145
This function can be alternatively considered as a function on the group C4 x Q2.The Fourier spectrum o f f is thus given by
G32 =
Sf
=
1 -[19, -3, -1,1,0, i , - i , i , 32
-2,
R, 1,-1, - 3 , 3 , V, -2, i , -2, i, -R]'r,
where
o=
-4
-
42
6i
v
42 -4+6i = -4-2i -8-4i
[
. I
-2
8-42
-4+2i
+ 4i
1.
0
J
Therefore, this function can be represented by the matrix-valued FNADD shown in Figure 4.47. In this decision diagram, the matrix-valued nodes are transferred into functions on C4 by using the algorithm A2 and represented by the nodes S, with four outgoing edges. Figure 4.48 shows the thus derived complex-valued FNADD o f f . Thisfunction f can be transferred into the matrix-valuedfunction f m on Q 2 given by the vector [F7rJ
=
[[
0 1 1 0 1 0 1 4 1 1 ]
3
[
0 1 0 1 1 0 1 4 1 01;
The Fourier spectrum of f i n is giveri by
where
Sf(1) = S f ( 2 ) = S f ( 3 ) = 8
-:I:
[ -:
146
OPTIMIZATION OF DECISION DIAGRAMS
Table 4.14 Complexities of representing f in Example 4.15 in terms of the number of levels, non-terminal nodes (ntn), constant nodes, (cn), and sizes.
Number of
d
ntn
cn
S
BDD
6
9
2
11
3
4
6
10
4
7
8
15 -
Matrix-valued FNADD Complex-valued FNADD FNAPDD
4 -
Complex-valued 15
1
Sf(4) = 4
It follows that f can be represented by the matrix-valued FNAPDD in Figure 4.49. Figure 4.50 shows FNAPDD with elements of matrix-valued nodes represented by matrix-valued MTDDs. If these elements are represented by decision diagrams, the function f can be represented by the complex-valued FNAPDD in Figure 4.51. Complex@ of representation of the considered function by BDD, FNADD and FNAPDD is compared in Table 4.14.
In this example, FNAPDD permits reduction of the depth and the number of non-terminal nodes without exceeding the size of BDD for f . This reduction is obtained as a trade-off between the complexity of the overall decision diagram and the complexities of the nodes of the decision diagram. The following example illustrates construction of FNAPDDs over finite fields.
Example 4.16 Figure 4.52 shows MTDD representation of the function f given by the following truth-table (takenfrom [14]).
CONSTRUCTION OF FNAPDD
Fig. 4.52 MTDD for f in Example 4.16 on GZ4= C2 x C2 x Cz x C 3 .
a=w,o+w,p+w,4
e=w,0+w2'+w,3+w,S
b=W,'+w;+w;
pw,)+w,4
C'W,O+W,'+W~ &W,s
fw,3
2
p2'+w,Z+w,4+w,S
h=w;+wl
Fig. 4.53 MTDD for f in Example 4.16 on G24 = C4 x C6.
147
148
OPTIMIZATION OF DECISION DIAGRAMS
Fig,4.54 mvFNAPDD for f in Example 4.16.
Fig. 4.55 FNAPDD for f in Example 4.16 with elements of mv nodes.
CONSTRUCTION OF FNAPDD
149
r
Fig. 4.56 ivFNAPDD for f in Example 4.16.
f 0
2122x324
f
x1x2x3xq
0100
1000
1
0101
1001
0
0002
1
0102
0 0 0
f 0
1002
0
0010
0
0110
0
1010
1
0011
0
0111
0
1011
1
1111
0012
I
0112
1
1012
0
1112
2122X3Xq
0000 000 1
1110
0 0
I f f is regarded as a function on G24 = C4 x (36, where c, and C6 are the groups orders 4 and 6, it can be represented by the MTDD with two-levels in Figure. 4.53. I f f is transferred into the matrix-valued function f m on the symmetric group of permutations S, over G F (11) by using the Algorithm 4.1, it is given by of
150
OPTIMIZATION OF DECISION DIAGRAMS
By using the Fourier transform for matrix-valued functions on Ss, the Fourier transform of fm on S, is given by
[Sf 1
where
9 - [ l o
S f ( 2 )=
-
10 21
[::]
2 [lo 2 [ I
10
6 1 10
5 1 -
ALGORITHM FOR CONSTRUCTION OF FNAPDD
~
LUT,
I
FFT
-
/”
p-
\
151
f
-
Fig. 4.57 Circuit realization of functions from FNAPDDs.
represented. Figure 4.56 shows this FANPDD with matrix-valued nodes represented by decision diagrams, which are reduced to a node with four outgoing edgesf o r each submatrix. In this way, integer-valued FNAPDDs (ivFNAPDDs) are derived. Complexities of representing f by using MTDD f o r G2j = C2 x C2 x C2 x C,, arid G24 = C, x CGand FNAPDD is compared in Table 4.15. Examples 4.14,4.15, and 4.16 are all considering binary-valued functions where BDDs and MTDDs have a starting advantage of 2 constant nodes, since they are defined with respect to the trivial basis. In the general case, for arbitrary discrete functions the number of different Fourier coefficients is not greater than the number of different values a function can take. That means, FNADDs and FNAPDDs may have smaller number of constant nodes that MTDDs. That especially applies to the functions with so-called flat spectrum. FNAPDDs permits to represent a function through matrix-valued coefficients that can be recorded in a Look-up-table. It follows that FFT processors can be used to calculate the function values from spectral coefficients. The same approach can be used for circuit realizations of discrete functions. Figure 4.57 shows the basic principle of realization of functions through FNAPDDs. 4.13
ALGORITHM FOR CONSTRUCTION OF FNAPDD
In practical applications of FNAPDDs representations, we assume that a library of finite Abelian and non-Abelian groups and their representations is provided. The groups whose orders are compatible to the class of functions that can be represented are taken in this library. The orders of the selected groups and their representations should be factors in the orders of vectors representing the targeted class of functions. For example, the quaternion group Q 2 is convenient in representation of switching
152
OPTIMIZATION OF DECISION DIAGRAMS
Table 4.75 Complexities of representing f in Example 4.16 in terms of the number of levels,non-terminalnodes (ntn), constant nodes (cn), and sizes.
Number of
G24 G24
=
c,x c,x c,x c,
= C, x
d
ntn
cn
s
5
7
2
9
3
5
2
7
2
1
3
4
C,
FNAPDD Number-valued FNAPDD
4
8
8
1
6
functions, since has the order 8 = 2 , and the representations of order 2. Similar, the symmetric group of permutations 5’3 is suitable for representation of three-valued and four-valued functions, since it is of order 6 with representations of order 2. 4.13.1
Algorithm for representation
For agiven f ,the FNAPDD representation is derived by using the following algorithm. Algorithm 4.5 (Determination of FNAPDD) 1. Given a function f t P(G). Determine a non-Abelian group G, of order
gn 5 g with representations of order r , by using the Algorithm 6.
2. Transfer f into the corresponding matrix-valued function f T n by using the Algorithm Apum. f, is represented by a vector [F,] of order y, whose elements are ( r , x r,) matrices. 3. Represent f, by the matrix-valued FNAPDD on G,.
4. Represent the matrix-valued nodes in FNAPDD of .fm by DDs on groups of order r;. 5. Do the possible optimizations of this representation.
Algorithm 4.6 (Generation of GJ
I . Assume that f may be represented by a vector G of order g 2. Read the orders of available groups gr and their representations Itwlin librury of jinite groups. Factorize g such that at least one factor of the form r i , g t appea rs.
OPTIMIZATION OF FNAPDD
153
3. Choose the corresponding non-Abelian group of order g , with representations of order rwgas a possible constituent subgroup G, in G,.
4. Repeat 2 $possible, otherwise choose the remaining constituent subgroups in G , among the available Abelian groups. In that, do a reasonable compromise between the number of levels and number of outgoing edges of nodes associated to the chosen subgroups G, in G,. 5. Determine the structure of G, as a group of orderg, = $ with representations of order r,, = k r w z ,where k is determined in previous step and order the subgroups G, chosen in 3 and 4 in the increasing order:
n,=,
6. Determine the Fourier transform matrix R on G, as the Kronecker product of the Fourier transform matrices on the constituent subgroups G,. 4.14
OPTIMIZATION OF FNAPDD
There are several possibilities for the optimization of FNAPDDs.
1. Choose the best suited group G, depending on the possible factorizations of the order g of G on which f is initially given. 2. Do the optimization in generation of f m from f by choosing between algorithms A1 and Az.
3. Do the optimization by choosing among DTs of various structures for the representation of the matrix-valued nodes in the matrix-valued FNAPDD. DDs on both Abelian or non-Abelian groups may be used. 4. Do the optimization in transferring the matrix-valued nodes into functions on the corresponding subgroups by choosing arbitrarily between the algorithms Aarnv by columns and by rows.
5. In representation of multi-output functions shared FNAPDDs may be used. These DDs are defined in the same way as Shared binary decision diagrams (SBDDs) and Shared multi-terminal binary decision diagrams (SM'IBDDs) are defined [I].The optimization may be done as with SBDDs and SMTBDDs by ordering of the variables in generation the integer-valued counterpart f ( z ) and by pairing of outputs.
154
OPTIMIZATION OF DECISION DIAGRAMS
REFERENCES 1 . Babu, H.Md.H., Sasao, T., “A method to represent multiple-output switching functions by using binary decision diagrams”, SASIM1’96, November 25, 1996.
2. Bernd, J., Gergov, J., Meinel, C., Slobodova, A,, “Boolean manipulation with free BDDs, first experimental results”, Proc. European Design and Test Con$, February 28 - March 3, 1994,200-207. 3. Bryant, R.E. “Graph-based algorithms for Boolean functions manipulation”, lEEE Trans. Comput., Vol. C-35, No. 8, 1986, 667-691. 4. Bryant, R.E., Chen, Y-A., “Verification of arithmetic functions with binary moment decision diagrams”, May 31, 1994, CMU-CS-94-160. 5. Clarke, E.M., McMillan, K.L., Zhao, X., Fujita, M., “Spectral transforms for extremely large Boolean functions”, in Kebschull, U., Schubert, E., Rosenstiel, W., Eds., Proc. IFIP WG 10.5 Workshop on Applications of the Reed-Muller Expansion in Circuit Design, 16-17.9.1993, Hamburg, Germany, 86-90.
6. Drechsler, R., Becker, B., OKFDDs-Algorithms, applications and extensions”, in Sasao, T., Fujita, M., (Ed.), Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996, 163-190. 7. Falkowski, B.J., Rahardja, S., “Complex spectral decision diagrams”, Proc. 26th Int. Synip. on Multiple-Valued Logic, Santiago de Campostela, Spain, May 1996, 255-260.
8. Hengster, H., Drechsler, R., Eckrich, S., Pfeiffer, T., Becker, B., “AND/EXOR based synthesis of testable KFDD-circuits with small depth”, Proc. Asian Test Corg, 1996. 9. Karpovsky, M.G., “Fast Fourier transforms on finite non-Abelian groups”, IEEE Trans. Comput., Vol. C-26, No. 10, Oct. 1977, 1028-1030. 10. Karpovsky, M.G., Trachtenberg, E.A., “Some optimization problems for convolution systems over finite groups”, Inform. Control., Vol. 34, No. 3, 1977, 1-22.
1 I . Karpovsky, M.G., Trachtenberg, E.A., “Fourier transform over finite groups for error detection and error correction in computation channels”, Inform. Control, Vol. 40, NO. 3, 1979, 335-358. 12. RoLiner, T.D., Karpovsky, M.G., Trachtenbcrg, L.A., “Fast Fourier transforms over finite groups by multiprocessor system”, IEEE Trans., Vol. ASSP-38, No. 2, 1990,226-240.
13. Lai, Y.F., Pedrarn, M., Vrudhula, S.B.K., “EVBDD-based algorithms for integer linear programming, spectral transformation, and functional decomposition”,
OPTIMIZATION OF FNAPDD
155
IEEE Trans. Computer-Aided Design of Integrated Circuits and Systems, Vol. 13, NO. 8, 1994, 959-975. 14. Sasao, T., Logic Design: Switching Circuit Theory, Kindai Kaga-ku, Tokyo, Japan, 1995. 15. Sasao, T., Switching Theory for Logic Synthesis, Kluwer Academic Publishers,
Boston, 1999.
16. Sasao, T., “Representations of logic functions by using EXOR operators”, in Sasao, T., Fujita, M., (eds.), Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996, 29-54. 17. Sasao, T., Fujita, M., Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996. 18. Sasao, T., Butler, J.T., “A design method for look-up table type FPGA by pseudoKronecker expansions”, Proc. 24th Int. Symp. on Multiple-valued Logic, Boston, May 25-27, 1994,97-104. 19. Sasao, T., Butler, T.J., “A method to represent multiple-output functions by using multi-valued decision diagrams”, Proc. Int. Symp. on Multiple-Valued Logic, May 29-3 1, 1996, Santiago de Campostela, Spain, 248-254. 20. Shannon, E.C., ”The synthesis of two-terminal switching circuits”, Bell System Tech. J., Vol. 28, No. 1, 1949. 21. StankoviC, R.S., “Some remarks about spectral transform interpretation of MTBDDs and EVBDDs”, Proc. ASP-DAC’95, August 29 - September 1, 1995, Makuhari Messe, Chiba, Japan, 385-390. 22. StankoviC, R.S., “Fourier decision diagrams for optimization of decision diagrams representations of discrete functions”, Proc. Workshop on Post BinaryUltra Large Scale Integration, Santiago de Campostela, Spain, 1996, 8- 12. 23. StankoviC, R.S., “Simple theory of decision diagrams for representation of discrete functions“, Proc. Reed-Muller 2000, Victoria, BC, Canada, May 20-21, 1999, 24. StankoviC, R.S., Astola, J.T., “Design of decision diagrams with increased functionality of nodes through group theory“, IEICE Trans. Fundamentals, Vol. E86-A, NO. 3,2003,693-703. 25. StankoviC, R.S., Astola, J.T., Spectral Interpretation of Decision Diagrams, Springer, New York, 2003. 26. StankoviC, R.S., Astola, J.T., “Matrix-valued decision diagrams in representations of complex systems“, 17th European Meeting on Cybernetics and Systems Research, Vienna, Austria, April 13-16, 2004.
156
OPTIMIZATION OF DECISION DIAGRAMS
27. StankoviC, R.S., Sasao, T., Moraga, C . , “Spectral transform decision diagrams”, in Sasao, T., Fujita, M., (eds.), Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996,55-92.
Functional Expressions on Quaternion Groups In this chapter, we discuss arithmetic expressions on the quaternion groups [ 191 corresponding to the arithmetic expressions on C; . The interest in arithmetic expressions on Q; is motivated by the renewed recent interest in arithmetic expressions on CX; originating in their properties and applications
1. Parallelization of algorithms for large switching functions [9], [ 101; 2. Single expression for multi-output switching functions [ 141;
3. Some word-level decision diagrams are graphic representations of arithmetic expressions for discrete functions [ 171, [ I 81, 1201; 4. Different applications of arithmetic expressions such, as for example, checking of error probability in logic networks [9], [22]. We extended a method for derivation of polynomial expressions from Walsh (Fourier) expansions on CX;to non-Abelian groups by using as the example the group (n-1)/3 2 (n--2)/3 Q 2 . For other values of n, we use the groups QT’3, C2Q2 , and C2Q2 . In this case, the arithmetic expressions are generated by the Kronecker product of the arithmetic matrices on C2 and Q 2 . We consider groups of orders 2n, therefore, it is possible to derive arithmetic expressions on quaternion groups in terms of switching variables. Extension to groups of arbitrary orders is possible by allowing variables to take values in arbitrary finite sets. An example of such generalizations for groups Cpn, for example p = 3, is given in [ 121. 157
158
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
5.1 FOURIER EXPRESSIONS ON FINITE DYADIC GROUPS 5.1.1
Finite dyadic groups
Group representations for Cz are given by the basic Walsh matrix
W(l)=
[
1 -11 1 .
The group representations of the finite dyadic group Cg are given by the Walsh matrix of order n defined as
where @ denotes the Kronecker product. We denote by C(C;) the space of functions f : Cg C , where C2 is the cyclic group of order 2 , and C is the complex-field. Assume that f E C(C;) is given by the vector F = [f(O), . . . , f (2n - 1)IT. If the Walsh (Fourier) spectrum for f is represented by a vector of spectral coefficients S,,f = [S,,f(O),. . . , S,,f(2n - 1)IT,then ---f
S,,f
= 2-nW(n)F,
and
since the Walsh matrix is a self-inverse matrix over C with the scaling factor 2-".
Example 5.1 The Walsh spectrum ofa three-variable switchingfunction f ( 5 1 , x2,x3) given by the truth-vector F = [I,O,O, O,O, 1,1,1ITis given by vector of coe#cients S,,f = [ 4 , 0 , 0 , 0 ,-2,2,2,2]'. Thus,
where w,(x)represent the columns of the corresponding Walsh matrix. 5.2 FOURIER EXPRESSIONS ON Q2 The Fourier transform on Q 2 is defined in terms of the group representations given by the columns of the matrix whose entries are determined from Table 2.9. From this table, the set offunctions R$')(x) used to define the Fourier transform on Q 2 is
FOURIER EXPRESSIONS ON Q 2
159
represented by the columns of the following matrix
Q2 =
1 1 1 1 1 1 1
1 1 1 1 0 0 1 1 -1 -1 i 0 0 -2 1 1 1 - 1 0 0 - 1 1 - 1 - 1 - i 0 0 i - 1 1 - 1 0 - 1 1 0 -1 -1 1 0 -2 -2 0 - 1 1 - 1 0 1 - 1 0 1 - 1 - 1 1 0 i i 0
Notice that in this matrix, the order of group representations is a bit different compared to Table 2.9, in particular, R1 and R2 are permuted. The Fourier transform matrix is given by the matrix inverse to Q 2 . Thus, 1 1 1 1 1 1 1 1 1 1 1 1 -1 -1 -1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 -1 1 -1 1 2 -2i -2 22 0 0 0 0 0 0 0 0 -2 22 2 -22. 0 0 0 0 2 2i -2 -2i 2 2i -2 -2i 0 0 0 0
We denoted by C(Q2)the space of functions f : Q 2 For a function f E C(Q2)given by the vector
F
=
---f
C.
[f(01,f ( l )f,( 2 ) , f ( 3 ) , f (4), f (51, f (61, f ( 7 ) I T ,
the Fourier spectrum given in the matrix notation by the vector of spectral coefficients S,.f = [S,.f(O), S,,f(l)>S,.f(2), S,,.f(3),S,,f(4), S,,f(5),S,,f(G),Sq.f(7)IT,
is determined as
and
F
= Q2Sq,f.
Example 5.2 I f f in Example 5.1 is considered as a function on Q 2 , then the Fourier spectrum f o r f is given by S , , f = i [ 4 , - 2 , 0 , 2 , 2 , 2 - 2,2]'. Thus, 1
f ( x ) = p ( 4 q 0 ( ~-) 2ql(z)
+ 2q3(z) + 2q4(x) + 2 q d z )
where qi (x)are columns of the matrix
Q2.
-
%(x)
+ 2q7(x)),
760
FUNCTIONAL EXfRfSSlONS ON QUATERNlON GROUPS
5.3 ARITHMETIC EXPRESSIONS In the matrix notation, the arithmetic expressions for functions in C(CF) are defined as
with
F
=
[ f ( 0 ) ., . . , f ( 2 “
-
l)IT,and n
where A( 1)is the basic arithmetic transform matrix given by
If the elements of X, are interpreted as logic values 0 and 1, then this matrix is referred to as the Reed-Muller matrix or the conjunctive transform matrix [2],[3]. It defines a self-inverse transform in GFz(CF)denoted as the Reed-Muller transform, or the conjunctive transform. In this context, the arithmetic transform in C(C2) is denoted as the inverse conjunctive transform [2]. For more details on arithmetic expressions, see for example, [2],[91, [lo], [151, [221.
Example 5.3 For 11 = 3, the arithmetic transform in C(C,”)is defined by the matrix
A(3) =
1 -1 -1 1 -1 1 1 1 -1
0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 - 1 - 1 1 0 0 0 0 0 0 1 0 0 - 1 0 0 - 1 1 0 - 1 0 0 - 1 1 0 0 - 1 0 - 1 0 1 1 1 -1 1 -1 -1
For f in Example 5. I , the arithmetic spectrum is given by Af = [l,-1, -1,1, -1,2,2, -21
Therefore,
T
.
0 0 0 0 0 0 0 0 1
ARITHMETIC EXPRESSIONS FROM WALSH EXPANSIONS
161
5.4 ARITHMETIC EXPRESSIONS FROM WALSH EXPANSIONS
Let G = C; and P = C. Each x E G can be expressed by x = ( X I ,x2),X I ,x2 E (0,l). Fourier expansions for functions f E C(C,”)are defined in terms of Walsh functions given by the columns of Walsh matrix
The set of Walsh functions wz,i ing variables as
= 0,1,2,3can be represented
in terms of switch-
= 1,
1.
wg
2.
w1 =
3.
w 2
= 1 - 2x1,
4.
1113
=
1 - 2x2,
(1 - 221)(1
~
2x2).
In symbolic notation, W ( 2 ) can be written as
x(2)=
[
‘(110
7111
1U2
2113
].
From (5.2), the Walsh expression for ,f E C(C;)is defined by
f
= =
X(2)SWl.f [ 1 1 - 2x2
1 - 2x1
(1 - 2x1)(1-222)
] Sf,
where S , , f = [S,,f(O), SW,f(l), S , J ( ~ ) S, w , , f ( 3 ) ]isTthe vector of Walsh spectral coefficients. With this notation, the orthogonal Walsh series expression transfers into the Walsh polynomial expressions in terms of switching variables. It is assumed that switching variables are coded by ( O , l ) c ; ~ ( --+ ~ ) (0,l)z. Therefore, f
=
1’ S,JjO)
+ (1
+(1 - 221)(l
-
~
2Z2)SW,f(l) + (1 - 2Zl)SW,f(2)
222)SW,f(3).
(5.3)
In (5.3), i f the multiplications are performed, then the polynomial expression for f is derived. f =I
. Zg
~
22122 - 2 X 2 Z 1 f 4X12223.
where 20
=
z1
=
22
=
+ S,,f(l) + S,,.f(2) + %,I.f(3), + SW.f(3), SWJ(2) + SW.f(3),
ZC(
=
S W J (3).
S,.f(O) S,,f(l)
(5.4)
162
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
Note that in this relation, indices of variables and coefficients are ordered in a way that corresponds to the Hadamard ordering of Walsh functions [8]. If Walsh (Fourier) coefficients are expressed in terms of the function values for f , these polynomial representations become the arithmetic expressions for f . In this example, SW,f(O)
=
1 ,(f(O)
S W , f P )
=
,(f(O
-f(1)+f ( 2 )- f(3)),
SUJ.f(2) =
1 ,(f(O)
+f ( 1 )- f ( 2 )
5L,f(3)
,(f(O
=
1
+ f ( 1 )+ f ( 2 )+ f ( 3 ) ) .
1
-
-
f(3)),
f ( 1 )- f ( 2 ) + f(3)).
Therefore,
After replacement of z, in (5.4), we get the arithmetic expressions for f E C(C,”)
f
= 1 . a0
+ x1u2 + X Z U I + x1x2a3,
where
If further
1 . Binary values 0 and 1 for variables are considered as the logic values 0 and I , 2. The addition and subtraction in C are replaced by the addition in G F ( 2 ) ,
3. Values of coefficients are calculated modulo 2, then, the arithmetic expressions become the Reed-Muller expressions for f , In this example, the Reed-Muller expression is given by
ARITHMETIC EXPRESSIONS ON Q2
163
where ~i E ( 0 , l}, are To
=
TI
=
1-2
= =
7-3
Example 5.4 For represented as
ri =
f (O), f (0) @ f ( l ) , f (0) @ f (4, f ( 0 ) @ f ( 1 ) 63 f (2)
@
f
(3)
3, the Walsh functions in the Hadamard ordering can be
r f we replace these relations in (5.1J, then
which is the arithmetic expression f o r f . If we reduce the coeficierirs modulo 2, and replace the addition and subtraction by EXOR, we get the Reed-Muller expression for f
f 5.5
= 1@ 2 1
(ij 2 2
a 2 3 @ 22x3.
ARITHMETIC EXPRESSIONS ON Qz
Let G be the Quaternion (non-Abelian) group Q2. The columns of the matrix Q 2 in terms of which the Fourier transform on Q 2 is defined can be represented by the following set of functions in terms of switching variables
1. qo
=
1,
2. q1 = (1- 22,),
3. q2 = (1 - 2x3), 4. q3 = ( 1 - 22,)(1
-
253),
164
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
5.
q4
+ 253), 252)(23 + Z Z ~ ) ,
= Zl(1 - 252)(s3
6. q5 =
- ~ 1 ( 1-
7.
q6
= x1(1 - 2 2 2 ) ( 3 3 - 2 x 3 ) ,
8.
q7
=Zl(l
-
222)(33 - 2 x 3 ) .
Each f E C(Q2)can be represented as a linear combination of these functions. The coefficients in this expression are the Fourier coefficients on Q 2 over C. Expressions,Arithmetic on non-Abelian groups Example 5.5 Consider a three-variable switchingfunction f as a function on Q2. If f is given by the truth-vector F = [l,O , O , O,O, 1,1,1IT,then the Fourier coeficients for f are given by the vector Sf = ;[4, -2,0,2,2,2, -2,2]’. Therefore, f
=
+ +
1 -(4 - 2(1 - 221) 2(1 - 2 5 l ) ( l - 252) 8 +2Zl(1 - 2rC2)(33 iz:3)- 221(1 - 2zz)(c3
+ 223)
-251(1 - 2 2 2 ) ( : 3
-
-
2x3)
+ 221(1 - 222)(T3
253)).
Finally,
+ 85153 + 4213
1 f = -(4 - 453 8
-
451%3 - 8z122z3
+ 821X2c3).
(5.5)
I n a general case,
f
+ 41Sq,f(l) + 4 2 S q A 2 ) + Q3Sq.f(3) +44sq,f(4) + 4 5 s q . f ( 5 ) + 4 6 s q . f ( 6 ) + q7sq.f(7).
YOSq,f(O)
=
From the matrix of the group representations on Q 2 , the Fourier coefficients on are given by Sq,f(O) = Sq,f(l)
=
1
,(f(O) + f(1) + f ( 2 ) + f ( 3 ) + f ( 4 ) + f ( 5 ) + f ( G ) + f ( 7 ) ) 1
,(f(O) 1
S q J ( 2 ) = ,(f(O) 1
f
f(1)+f ( 2 )+f(3)
- f(1)
-
f ( 4 ) - f ( 5 ) - f ( G ) - .f(7))
+ f(2) - f ( 3 ) + f(4) - f ( 5 ) + f(6)- f(7))
- f(1)+ f ( 2 ) - f ( 3 ) - f ( 4 ) + f ( 5 )
Sq..f(3)
=
,(f(O)
SqJ(4)
=
p o l
sq.f(5) =
1 -(-2f(4) 8
Sq,f(G) =
-(2f(4)
Sq,f(7) =
Q2
1
1
8 1 -(2f(0) 8
-
2if(l)
-
2”f(2) + 2 i f ( 3 ) )
+ 2if(s) +2f(6) + 2if(5) + Zif(1)
-
aif(7))
2if(7))
-
2f(6)
-
2f(2) - 22f(3))
-
-
f ( 6 )+ f ( 7 ) )
ARITHMETIC EXPRESSIONS ON Q 2
165
From there.
x, =
1 1 1 1 1 1 1 1
0 1 0 1 1 0 0 1
0 0 0 0 1 1 1 1
0 0
1 0 0 0 1 0 0 - 1 0 0 0 0 - 1 0 1 0 0 1 0 0 0 0 0 0 0 - 1 1 0 0 0 -
-
0 0 0 0 0 . 1 0 1 -
r f f E C(Q2)is given by a vector F = ( f ( 0 ) ., . . , f (7)IT, then the coeficients Q f= [qo, . . . , q7IT in the expansion for f with respect to the basis given by X, are
determined as
Q f= X i l F , where Xp' is the matrix inverse for X, over C. Thus, 1 0 1 0 0 -1 1 - 1 1 0 -1 0 - 1 0 1 1 -1 1 -1 -1 1 0 - 1 0 0 0 1 0 - 1 0 0 0 0 0 1 0 0 0 0 0
0 0 0 0 0 1 1 -1 0 0 0 0 0 - 1 1 0 -
0 0 0 1 0 ' 0 0 1 -
This matrix can be considered as a transform matrix de$ned with respect to the basic functions represented by columns of the matrix X,. Thus, this matrix defines the inverse transform. We call this transform the Arithmetic-Haar transform, since the matrix X i 1 exhibits a form similar to that of the H a m transform matrix [8].
166
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
Example 5.6 For f in Example 5.1, the arithmetic-Haar coeficients are given by the vector sf = $11, -1,0, 2 , 1 , 0 , -l,oIT. Therefore,
Extensions of the arithmetic-Haar expressions from Q 2 to Qg, 'n = 3r, can be done through the Kronecker product of the basic transform matrices on Qz in the same way as in the case of the arithmetic expressions and Walsh series expansions. Generalization to other values for n is easy, by using combinations of the basic arithmetic transform matrices. For example, if for a four variable switching function, the domain group is chosen as G = C2 x Q2, then the arithmetic-Haar transform matrix is defined as the Kronecker product of A ( l ) and X,.
5.5.1 Arithmetic expressions and arithmetic-Haar expressions In derivation of arithmetic-Haar expressions, we have exploited the correspondence between the order of C; and Qy'3. Due to that, we express the basic functions in arithmetic-Haar expressions in terms of switching variables, similar as that is done for the arithmetic expression on Cg. Therefore, it is natural to establish a relationship between the arithmetic expressions on Ci and the arithmetic-Haar expressions on Q;l3. The arithmetic-Haar expression for a given function f , converts into the arithmetic expression for f if we replace Z, -+ (1 - xz),and then recalculate the coefficients and assign them to the basic functions used in the arithmetic expressions.
Example 5.7 For f in Example 5.6, the arithmetic-Haar expression is converted into the arithmetic expression as follows:
Arithmetic-Haar expressions and Kronecker expressions If we use a different representations for functions 40, . . . , q 7 , we will derive different arithmetic expressions for functions in C(Q2).We use the above given representations, since they
1 . Resemble representations of Walsh functions in terms of switching variables,
2. Provide the transform matrix similar to the Haar transform.
DIFFERENT POLARITY POLYNOMIAL EXPRESSIONS
167
Arithmetic expressions in C(C,”)are series expansions in terms of basic functions generated as
If the values for variables are considered as the logic values, then the same basic functions are used in the definition of the Reed-Muller expressions. These basic functions can be generated by the recursive application of the positive Davio (pD) expansion defined as f = 1 . fo 8 z,(fo f l ) , where f o = f ( x z = 0), and f l = f ( 2 %= 1). Arithmetic-Haar expressions are series expansions in terms of the basic functions given by the symbolic matrix xq
[1 z1?&(1
=
23
- 222)
21x3
21
z123(1
-
252)
qz’3(1-222) 2123(1- 2 4 1 .
These basic functions can be considered as a combination of two sets of basic functions in C(C:). First four basic functions are generated by the pD-expansion for variables x1 and 5 3 . The other four basic functions are generated by the Shannon (S) expansion defined by f = Z,fo 69 zzfl for z1 and 2 3 , with multiplication by (1 - 252). Thus, we generate the first four basic functions in the arithmetic-Haar expressions as [ I
.1]@[1
2 3 ] = [ 1
23
21
21x31.
For other four basic functions we first use the S-expansion and get
[
21
21
]@[
23
23
]
[
21x3
5123
1.
Then, we multiply the generated terms by (1 - 222). For functions in GF2 (C;),expressions generated by combination of pD-expansions and S-expansions are denoted as Kronecker expressions [ 131. In this setting, the arithmetic-Haar expressions are modified integer counterparts of the Kronecker expressions for switching functions. 5.6
DIFFERENT POLARITY POLYNOMIAL EXPRESSIONS
Switching variables take values in Ca. Negative literals of switching variables are defined as E = 2 @ 1. The fixed-polarity Reed-Muller expressions are defined by freely choosing between z and zfor each variablc in f [ 141. It is assumed that a particular variable appears as the positive or the negative literal, but not both in the same expressions for f . For a given f , fixed-polarity Reed-Muller expressions differ in the number of non-zero coefficients. The polarity of variables chosen in an expression is conveniently expressed by the polarity vector H = ( h l , . . . , hn),h, E (0, I}. It is assumed that if h, = 1,then the
168
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
z.
i-th variable is represented by the negated literal Thus, h, = 0, shows that the i-th variable appears as z,. The polarity where h, = 0 for each i = 1, . . . , n is denoted as the zero-polarity. For a given class of expressions and a given f , the polarity for which the expression has the minimal number of non-zero coefficients is denoted as the optimal polarity. In Fourier expansions for functions in C(Q2),the basic functions can be expressed in terms of switching variables. Thanks to that, the same method can be used to define fixed-polarity Fourier expansions C(Qa),and the arithmetic-Haar expression (n-1)/3 2 (n-2)/3 for functions in C(C,")and C(Q;/'), C2Q2 , and C2Q2 . 5.6.1
Fixed-polarity Fourier expansions in C (Q2)
Fixed-polarity Fourier expansions on Q 2 will be introduced through the example for H = (101).
Example 5.8 The use of negated literals for variables X I and futictions for the Fourier transform on Q2 into
1. q
p = 1,
2. q
y = 1 - 221,
4. qhl"l) = (1 - 2:1)(1
6. q k l O 1 ) = -Z1(1
7. q
-
-
-
transfers the husic
223).
2x2)(53
p = :rl(l - 2 2 2 ) ( 2 3
8. q i l O 1 ) = D', (1
23
252)(53
+ E3),
-
izy),
-
2.3).
Thus, the matrix ofbasic,functions Qzfor the Fourier transform in C(Q2)for the polaritj H = (101) is given by
r l 1 1 1 1
-1 - 1 - 1 -1 = 1 1 1 1 1 1 1 -
QFol)
-1 1 1 1 -1 1 -1 1
1 0 7 -2 - 1 0 1 1 1 0 - 2 7 -1 0 -1 -1 -1 7 0 0 1 1 0 0 -1 - 1 0 0 1 -1 0 0
0 0 0 0 --I
1 2
-1
DIFFERENT POLARITY POLYNOMIAL EXPRESSIONS
169
It follows that the colunins of QFol) are given by qo, -ql, - q 2 , q3, iq6,iq7, -iq4,iq5. This permutation and change of columns in Qgol)requires apermutatiori arid charzge of rows in QY1. Wefirst transform rows in QT1 in the same way as we transform the columns in Qz to get Qgol).In this way we get
(Q8’)’
1 -1 -1 1 0 2i -2i 0
=
1 1 1 1 1 1 -1 -1 -1 1 1 1 1 -1 1 -1 1 -1 -1 1 -1 -1 1 -1 0 0 0 2i -2 -22 -2 -2i 2 0 0 0 0 2 0 0 2i -2 2i 0 -22 -2 0 0
-
(101) -1 ‘
)
(Qz
-
1 1 1 -1 -1 -1 -1 1 -1 1 1 -1 1 0 0 0 2 2i -2i 2 -2i 2i 0 0 0 -
-S
1 -1 1 -1
o
-2 -2 0
1 1 -1 -1 -2i 0 0 2i
-
1 1 1 1 2 0 0 2 -
1 1 1 1 1 1 1 -1 1 1 -1 1 2 2i -2 0 0 0 0 0 0 -2 2 -2i
-
-
1 SLtf”” = -[4,2,0.2,22, -2ii 22. -2iIT. 8
Therefore, 1
f = -(4 8
+ 2(1
-
2T1)
+ 2(1
-
-22T1(l - 252)(53 + iZ3)
2T1)(1 - 2Z3)
+ 2iz1(1 - 2x2)(x3 + iZ3)
+ 2iz1(1- 222)(x3 - iz3)
-2iXl(1 - 252)(X3 - ZZ3)).
5.6.2 Fixed-polarity arithmetic-Haar expressions
The matrix of basic arithmetic-Haar functions does not have complex values, and therefore, the transformation of this matrix into the matrix for the given polarity H is simpler than in the case of the Fourier transform. We will introduce the fixed-polarity arithmetic-Haar expressions in C(Q2)with the example for H = (101).
170
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
-
1 1 1 1 Xt01) = 1 1 1 - 1
0 1 1 0 1 0 1 0
1 1 1 1 0 0 0 0
0 0 0 0 1 1 0 0 1 0 1 0 0 0 - 1 0 0 0 - 1 0 0 0 1 0 0 0 1 0 0 0 0 0 - 1 0 0 0 - 1 0 0 0
0 0 0 0 0 1 0 0 0 0 0 1 - 1 1 0 1 0 1 0 - 1 0 1 -1 1 -1 -1 1 -1 0 0 0 0 0 1 0 0 0 0 0 1 0 - 1 0 1 0 - 1 0 0 0 1 0 - 1 0 0 0 0
1 1 1 1 1 0 0 0
For f in Example 5.1, the arithmetic-Haar expression for the polarity H = (101) is given by
This expression requires six coeficients instead offive coeficients in the zero-polarity arithmetic-Haur expression for f . Table 5.1 shows diferent spectra for f in Example 5.1, for the zero-polarity and the polarity H = (101). Table 5.2 shows the corresponding expressions for f . Table 5.3 compares the number of non-zero coefficients in some mcnc benchmark functions used in logic design for the Walsh (W), the arithmetic (A), the Fourier on Qz (F), and the arithmetic-Haar (AH) transform. For multiple-output functions, each output is considered as a separate switching function, and in this table, fun-i denotes the i-th output of f u n . This table shows that for each transform we can find a function where this transform provides a spectrum with fewer number of non-zero coefficients. In the most cases, the arithmetic and the Walsh transform are the most efficient, since are applied to binary-valued functions. However, even the complexvalued Fourier transform although applied to binary-valued functions can produce simpler spectrum than these transforms. For example, for sao2-2, i = 1 , 2 , 3 , 4 , the Fourier transform is more convenient than both the arithmetic and the Walsh
DIFFERENT POLARITY POLYNOMIAL EXPRESSIONS
Table 5.7
Different spectra for
f
SDectrum i1,1,1,1,1,0,0,0,IT i1,1,0,1,1,0,0, 1IT [l,-1, -1,1, -1,2,2, -2]T
[I,- l , O , 1, -1,2,0, -2]T
' 2[1, 1 -1,0,2,1,0,
-1,OlT
' 3 2 , -1, -2,2,0, - l , O ,
l]T
Table 5.2 Different expressions for f
171
172
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
transform. However, in this case, the arithmetic-Haar transform further reduces the number of non-zero coefficients. Conversely, for a given function, we should try different transforms to determine a spectrum with minimum number of non-zero coefficients. For example, for rd84-2 and rd84-3, the Fourier transform produces spectra with fewer non-zero coefficients than the arithmetic and the Walsh transform, respectively. For rd84-3, the arithmetic-Haar spectrum is comparable to the most efficient transform. For %ym, the arithmetic-Haar transform requires fewer nonzero coefficients than the arithmetic transform, and the Walsh transform is the most efficient. For Sxpl-2, the arithmetic-Haar transform is more efficient than the Walsh transform, and the arithmetic transform is the most efficient. However, for 5xp 1-3, the arithmetic-Haar transform is the most efficient. For all the outputs of sao2, the Fourier transform requires fewer products than either Walsh or the arithmetic transform, and the arithmetic-Haar transform is the most efficient. We consider this analysis as a justification for introduction and use of different transforms, supported by a requirement that these transforms may share some useful properties of existing transforms. In that respect, the arithmetic-Haar transform is derived from the Fourier transform on Qk in a uniform way as the arithmetic transform is derived from the Walsh transform on CF. It eliminates the complex values in the Fourier transform and the corresponding transform matrix expresses a structure similar to the structure of the Haar matrix.
Example 5.10 Assume the I6bits binary representation f o r integers x and y. Then, thefunction y = 21Gszn(i/216)can be consideredas 16-variable 16-output switching function f . y'we petjorm the addition of outputs multiplied with weighting coefficients 2', z = 0 , . . . 15, we determine an integer-valued equivalent function f i f o r f . Table 5.4 shows the number of non-zero coefficients f o r y in the Fourier spectrum and the arithmetic-Haar spectrum on C2Q5, and the Walsh spectrum, the arithmetic spectrum and the Huar spectrum on Ci6. It is interesting to note that the complexvalued Fourier spectrum requires the fewer number of non-zero coefficients. Among them 5248 are real, the 5120 are the imaginary, arid 7776 are the complex-valued coeficients.
.
5.7 5.7.1
CALCULATION OF THE ARITHMETIC-HAAR COEFFICIENTS FFT-like algorithm
The arithmetic-Haar transform matrix X, has a form similar to that of the Haar transform matrix. Thanks to that property, it is possible to derive a FFT-like algorithm for calculation of the arithmetic-Haar coefficients. This algorithm is based upon the
CALCULATION OF THE ARITHMETIC-HAAR COEFFICIENTS
Table 5.3 Number of non-zero coefficients.
f 5xpl-1
F A H
5xp 1-5 5~~1
96 122 124 70 84 - 6 64
26 50 34 56 36 44
5xp 1-7 5xpl-8
56 28
32 20
5xp 1-9 5xpl- 10
24 64
8 10
464 220 160 48 237
300 189 52 4 166
202 275 348 322
200 269 297 216
5xpl-2 5xpl-3 5xpl-4
9sym rd84- 1 rd84-2 rd84-3 rd84-4 sao2- I sao2-2 sao2-3 sao2-4
W
A
80 128
18
33
65 33 17 9
41 27 17 9
5 5 2 3 2 2 128 7 256 465 256 191 2 255 256 1 256 163 1024 380 640 768 856 578 1024 936
173
174
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
following factorization of X i 1
-
Xi’
z
1 0 1 0 0 0 0 0
0 1 0 1 0 0 0 0
1 0 1 0 0 0 0
0 1 0 0 1 0 1 0 0 - 1 0 0 1 0 - 1 0 0 0 0 1 0 0 0 0 0 0 1 0
0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0
-
1 0 - 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 - 1 1 0 -1 0 0 0 0 0 0 1 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 1 -1 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 1
-
.
Figure 5.1 compares the flow-graphs of the fast algorithms for calculation of the Walsh, the arithmetic, the Fourier, and the arithmetic-Haar coefficients for functions in f E C(Cz,and f E C ( Q 2 ) .
5.7.2 Calculation of arithmetic-Haar coefficients through decision diagrams Decision diagrams (DDs) are an efficient data structure for compact representation of discrete functions [I], [4], [ 141. Among many other applications, DDs are efficiently used to calculate different spectral transforms. From spectral interpretation of DDs [20], it was easy to show that in calculation of spectral transforms through DDs we actually perform basic operations in FFT-like algorithms. However. in DDs, calculation of the spectrum for f is not based on the vector of function values, but is performed over the DD for f . That permits to use the particular properties o f f , which provides the computation efficiency and possibility to process large functions [SI, 1141,1211. Calculation of the coefficients in arithmetic expressions of switching functions through DDs was considered in several papers, see for example 161, [7], [ 1 11. In this section, we present a method to calculate arithmetic-Haar coefficients through decision diagrams (DDs) [ 13, [4]. The method is derived as a suitable modification of the corresponding method for calculation of the Haar spectrum [ 161. We assume that a given f E C(Q5) is alternatively considered as a function in C(C,”),n = 3r, and represented by a Binary DD [ I ] , [4], or a Multi-terminal binary DD (MTBDD) [S],depending on the function values taken in G F ( 2 ) or C , respectively.
CALCULATION OF THE ARITHMETIC-HAAR COEFFICIENTS
Arithmetic
Walsh
Fourier
175
Arithmetic-Haar Fig. 5.1 FFT-like algorithms for 71 = 3 .
176
FUNC JlONAL EXPRESSlONS ON QUAJERNlON GROUPS
f i g . 5.2 BDT for f in Example 5.1. The method will be first explained on decision trees (DTs), since DDs are derived by the reduction of DTs [14].
Example 5.11 Figure 5.2 shows BDT f in Example 5.1. Figure 5.3 shows the corresponding DD. In this DD, two cross points [20] are shown.
Fig. 5.3 BDD for b' in Example 5.1. We will present the method in a general form for functions in C ( Q $ ) ,but we illustrate the method by the example of functions f E C(Q2).
CALCULATION OF THE ARITHMETIC-HAAR COEFFICIENTS
177
We assign to each node two fields, where we write the result of calculation through DDs. Calculations consist of processing the nodes in the DT for f . In the case of DDs, cross points [20] should be also processed. In a DD, the leftmost node at a level is the node which can be reached by the path labeled by a product of variables consisting of negated literals. In a DT, values of constant nodes are function values for f . We assume that at the level ( n - 1) in the DT for f , the paths T,-~:IL,~ and x,-~x, are permuted. Thus, in the vector F representing f ,the pairs of adjacent elements, starting from f (1) are permuted. For 7% = 3, f is given by F = [ f ( o ) , f ( l )f ,( 2 ) , f ( 3 ) , f ( 4 ) , f ( 5 ) , f ( G ) , f(7)I7’. Afier the mentioned reordering, the values of constant nodes in the DT for f are ordered as F, = f(21, f ( 4 ~ f ( 5 f(7)iT. ~ The nodes at the level X, are processed by using the rules described by rows of the matrix W(1). The values of the left and the right fields arc determined by using the first row and the second row of W ( 1 ) , respectively. The values of both fields for the nodes at all other levels are determined by using the rule described by the matrix [-1,1]. The input data for calculation of values of left and right fields are determined as follows. At the i-th level in the DT, ifthe node is pointed by the edge labeled with the negated literal T f ,the input data for the left and right fields are reached by the subpaths ?E,?E,+l and ? f 7 x L + 1respectively. , If the node is pointed by the edge labeled with the positive literal, then the input data for the left and right fields are reached by the subpaths x , Z , + ~and x,T,+.~,respectively. The coefficient S a , f ( 0 )is shown in the left field assigned to the leftmost nodc at the level x,. The values of left fields assigned to the other nodes at this levcl are used in further calculations. The right fields of nodes at the level :c, show values of arithmetic-Haar coefficients S,,J(~‘”-’)to Sa,f(2” 1). For n = 3, these are coeficients S,,f(4), S , , f ( 5 ) ,S,,f((i), and S,,f(7). The other coefficients are shown in the left and right fields of the leftmost nodes from the level 11. - 1 to the root node. These coefficients are shown in the increasing order. Thus, the leftmost node at the level :xTL-l shows ScL.f(l). while the right field of this node shows S,,f(2). The root node shows the coefficient S,,f(2’”-’ - 1). For [ I , = 3, the leftmost nodes at level 5‘2 shows the values ofS,,f(l)and S,.f(2). The root node shows S a , f ( 3 ) .
m), m, .m,fw,
~
Example 5.12 Figure 5.4 shows calculation qf urithmetic-Huur coeficientsfilr f t C (Q2) though 07:Figure 5.5 shows calculation qf arithmetic-Huar coeflcients for f in Exunzple 5.2 through BDT Figure 5.6 shows the same calculution through BDD. Note that in this example, perrriutution ofpaths ?E:zn::3and Z.LS:< dozs riotpermute values qfconstant nodes, since f(1)= , f ( 2 ) = 0, f ( 3 ) = f(4) = 0. and f ( 5 ) = f ( G ) = 1. This algorithm can be used to calculate arithmetic-Haar coefficients of complexvalued functions if instead of BDDs, we use MTBDDs.
178
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
Fig, 5.4 Calculation of arithmetic-Haarcoefficients through BDT
I
+l)
+ 1=2
SJ7) , 1 + 1 = 2 1-l=O; LX2
Fig. 5.5 Calculationof arithmetic-Haarcoefficientsfor f in Example 5.2 through BOT.
CALCULATION OF THE ARITHMETIC-HAAR COEFFICIENTS
179
Fig. 5.6 Calculation of arithmetic-Haarcoefficientsfor f in Example5.2 through BDD.
180
FUNCTIONAL EXPRESSIONS ON QUATERNION GROUPS
REFERENCES I . Ackers, S.B., “Binary decision diagrams”, IEEE Trans. Computers, Vol. C-27, No. 6, June 1978,509-516. 2. Agaian, S., Astola, J., Egiazarian, K., Binary Polynomial Transforms and Nonlinear Digital Filters, Marcel Dekker, New York, 1995.
3. Aizenberg, N.N., Trofimljuk, O.T., “Conjunctive transforms for discrete signals and their applications of tests and the detection of monotone functions”, Kibernetika, No. 5, (1981). 4. Bryant, R.E., “Graph-based algorithms for Boolean functions manipulation”, IEEE Trans. Computers, Vol. C-35, No. 8, 667-691, 1986.
5. Clarke, E, M., M.C., Millan, K.L., Zhao, X., Fujita, M., “Spectral transforms for extremely large Boolean functions”, in Kebschull, U., Schubert, E., Rosenstiel, W., Eds., Proc. IFlP WG 10.5 Workshop on Applications of the Reed-Muller Expressiori in Circuit Design, Hamburg, Germany, September 16-17, 1993, 8690. 6. Falkowski, B.J., Chang, C.H., “Efficient algorithm for the calculation of arithmetic spectrum from OBDD and synthesis of OBDD from arithmetic spectrum for incompletely specified Boolean functions”, Proc. 27th IEEE Int. Symp. on Circuits and Systems ISCAS94, London, Vol. 1, May 1994, 197-200. 7. Falkowski, B.J., Chang, C.H., “Calculation of arithmetic spectra from free binary decision diagrams”, Proc. IEEE Int. Symp. on Circuits and Systems (30th ISCAS), Hong Kong, June 1997, 1764-1767.
8. Karpovsky, M.G., Finite Orthogonal Series in the Design of Digital Devices, Wiley and JUP, New York and Jerusalem, 1976. 9. Kukharev, G.A., Shmerko, V.P., Yanushkievich, S.N., Technique of Bitzary Data Parallel Processing f o r VLSI, Vyshaja shcola, Minsk, Belarus, 199 1 . 10. Malyugin, V.D., Paralleled Calculation by Means of Arithmetic Polynonzials, Physical and Mathematical Publishing Company, Russian Academy of Science, Moscow, 1997. 1 I . Malyugin, V.D., StankoviC,R.S., StankoviC, M., “Calculations of the coefficients
of polynomial representations of switching functions through binary decision diagrams”, Proc. Preventive Engineering and Information Technologies, NiS, Yugoslavia, December 8- 12, 1994, 10-1- 10-4. 12. Moraga, C., “On some applications of the Chrestenson functions in logic design and data processing”, Mathenzutics and Computers in Simulation, 27, 1985,43I 439.
CALCULATION OF THE ARITHMETIC-HAAR COEFFICIENTS
187
13. Sasao, T., “Representations of logic functions by using EXOR operators”, in Sasao, T., Fujita, M., Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996, 29-54. 14. Sasao, T., Fujita, M., (eds.), Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996. 15. Shmerko, V.P., “Synthesis of arithmetical form of Boolean functions through the Fourier transform”, Automatics and Telemechanics, No. 5 , 1989, 134-142. 16. StankoviC, M., JankoviC, D., StankoviC, R.S., “Efficient algorithms for Haar spectrum calculation”, Scient$c Review, No. 2 1-22, 1996, 171- 182. 17. Stankovid, R.S., “Some remarks about spectral transform interpretation of MTBDDsand EVBDDs”, Proc. ASP-DAC’9.5,August 29-September 1,1995, Makuhari Messe, Chiba, Japan, 385-390. 18. StankoviC, R.S., Spectral Transform Decision Diagrams in Simple Questions and
Simple Answers, Nauka, Belgrade, 1998. 19. Stankovi c, R.S.,Moraga, C., Astola, J.T., “From Fourierexpansions toarithmeticHaar expressions on quaternion groups”, Applicable Algebra in Engineering, Communication and Computing, Vol. AAECC 12, 2001, 227-253. 20. StankoviC, R.S., Sasao, T., Moraga, C., “Spectral transform decision diagrams” in Sasao, T., Fujita, M., Representations of Discrete Functions, Kluwer Academic Publishers, Boston, 1996,55-92. 21. StankoviC, R.S., StankoviC, M., Jankovid, D., Spectral Transforms in Switching Theory, Definitions and Calculations, Nauka, Belgrade, 1998. 22. Yanushkevich, S.N., Logic Differential Calculus in Multi-Valued Logic Design, Techn. University of Szczecin Academic Publisher, Poland, 1998.
This Page Intentionally Left Blank
Gibbs Derivatives on Finite Groups Differential operators are a very powerful tool for the mathematical modeling of natural phenomena. Usually the differentiation is with respect to time or to a spatial coordinate, modeled by the real line R. In this setting, using the differential operators, the direction, as well as the rate, of change of a quantity can be adequately described. Moreover, by forming linear differential equations with constant coefficients, we get a very convenient way of expressing the principle of superposition inherent in many natural phenomena. Linearity offers an easily tractable model, usually sufficiently good as a first approximation. Fourier analysis, having linearity and the superposition principle in its essence, is another very efficient tool used for the same purposes. It is known from classical analysis that there is a strong relationship between the Newton-Leibniz derivative f ’ of a function f on R and its Fourier transform, which can be expressed by
where F’ and F denote the Fourier transforms of f‘ and f , respectively. Replacing the real group R by a locally compact Abelian, or acompact non-Abelian group extends classical Fourier analysis into abstract harmonic analysis. In this setting it has been natural to think about differentiation on groups in a way preserving as many as possible of the useful properties of Newton-Leibniz differentiation. The Gibbs derivatives on groups [4], [8], [ 101, [20], [241, [.331, [341, [401 form a class of differential operators extending the relation (6.1) to the functional spaces on other groups. In this more general context the role of the Euler functions exp(jtz) (the characters of the real group R ) is taken over by the characters of locally compact 183
184
GlBBS DERIVATIVES ON FINITE GROUPS
Abelian groups 141, 161, 181, [ 101, 1181, [20], or by the unitary irreducible representations of compact non-Abelian groups [22]. In this book attention is focused on Gibbs derivatives on finite, not necessarily Abelian, groups. We consider in detail one of the matrix representations suggested in [ I 11, which is suitable for the numerical evaluation of Gibbs derivatives of a given function. Using this matrix representation we present some FFT-like algorithms for calculation of the values of Gibbs derivatives on finite groups [29]. For a review of Gibbs differentiation see [30] and for some particular examples the bibliography 1121 given in 131. Some very recent results are reviewed in [30]and 1381, 1391. 6.1
DEFINITION AND PROPERTIES OF GIBBS DERIVATIVES ON FINITE NON-ABELIAN GROUPS
As is noted above, Gibbs differential operators on Abelian groups are defined as linear operators having the group characters as their eigenfunctions, see for example [lo]. Since the group characters are the kernels of Fourier transforms on locally compact Abelian groups, i t is very convenient to characterize the Gibbs derivatives by Fourier coefficients. Moreover, the strong relationship between the Gibbs derivatives and Fourier coefficients is somewhere used as the starting point for introduction of Gibbs derivatives on some particular groups, see for example 1181. By using the same approach the Gibbs derivatives on finite non-Abelian groups are defined in terms of Fourier coefficients as follows 1221.
Definition 6.1 Gihhs derivative D f of a furictiori f t P (G )whose Fourier trurisforni is Sf is defined by
As is noted in 1221 this definition is unique only by virtue of the fixed order adopted for the clcments of r. If a different notation was adopted, then (6.2), though unchanged in appearance, would define a distinct differentiator. This phenomenon is already present in the definition of the dyadic Gibbs derivative [7] which dcpcnds upon the order assumed for thc Walsh functions. The same applies to all other Gibbs derivatives on various groups. In what follows the Gibbs derivatives will be denoted by D f or, alternatively, by f(U.
An interpretation ofthe Gibbs derivative on finite non-Abelian groups can be given by following the approach used in [ 171 for the finite dyadic Gibbs derivative and later in 1.311 for the Gibbs derivatives o n finite Abelian groups. Define the partial sum f,(z).p 5 K by D- 1
(6.3)
DEFINITIONAND PROPERTIESOFGIBBS DERIVATIVES ON FINITE NUN-ABELIAN GROUPS
Define also the Fejkr sum as
c 4
adz) = 4-1
(6.4)
fP(.).
p= 1
Substituting (6.3) into (6.4) we have, after a simple calculation, K- 1
f ( z ) - CTK(Z) = K-'
C wT~(S~(W)R,(Z)).
w=o
The left member of this equality is the error in the approximation off by its Fejkr sum a ~ ( z )Thus, . the Gibbs derivative on a finite non-Abelian group G can be interpreted as that error multiplied by K . The chief properties of Gibbs derivatives are analogs to the corresponding properties of the classical Newton-Leibniz derivative, and they are given by the following theorem.
Theorem 6.1 I f f
E
P ( G ) ,then
I . D(aifi
+a2f2) = aiDfi +~ 2 D f 2W , , Q E~ P , f l , f 2 E P(G).
2. D f
t P iff f is a constant function.
=0
3. I f the Fourier transform off is Sf, then that o f f U I S ~w , = 0 , . . . , K 1.
( l ) is
given by Sf(" ( w ) =
~
This property can be interpreted as the fact that the set { R;") (x)} is the set eigenfunctiotis of the Gibbs derivative, i.e.,
of
From that, thanks to the linearin, of Gibbs derivatives,
D?'rR,(z) = wTrR,(x). 4. From the property 3, it easily follows that
D(fi
* f2)
=
(ofi)* f 2
=fl
* (Dfi),f1,f2
E P(G).
where * denotes the convolution on G
5. The Gibbs derivative commutes with the translation (shift) operator T on G dejned by ( T ' f ) ( z ) = f ( r o z-'), i.e., D ( T ' f ) = T ' ( D f ) , for each r t G . 6. It is known that the Gibhs differential operators do not obey the product rule. The same applies to the Gibbs derivatives on Jitiite non-Abelian groups, i.e., it
is false that for each f l and
f2
D(flf2) = f l ( D f 2 )+ (DfIlf2
185
786
GlBBS DERIVATIVES ON FINITE GROUPS
The Gibbs derivatives can be extended to an arbitrary complex order k by way of the definition of the delta function:
c
K-l
6(2)
=y
r,TrR,(x).
p
w=o
c
The &function thus defined has the property O(2) =
1, 2 = 0, 0, x # 0.
The Gibbs derivative of order k of the &function is obtained by a direct application of Property 3
c
K-1
6("(2)
wkr,,TrR,(x).
= 9-1
w=o
By using Property 4 of Theorem 2.2:
( D k f ) ( z= ) ( ( D ' h ) * f ) ( ~ =)
c
K-1
uPTTT(S,~(~)R~(~)).
w=o
6.2
GIBBS ANTI-DERIVATIVE
In this section we consider the determination of the values of a function from the values of its Gibbs derivative. It is obvious from (6.5) that the Gibbs derivative can be considered as a convolution operator on P ( G ) . More precisely, if we introduce a function W k defined by its Fourier coefficients as
where IT,,is the (ru, x r w )identity matrix, then from (6.5) the Gibbs derivative of order k of a function f E P ( G ) is given by ( D f ) ( w )= (tVk * f)(.).
From here we immediately deduce thc concept of the Gibbs anti-derivative. Introduce a function TV-k defined in the transform domain by
After the inverse Fourier transform CV-k(Z) =
1
+
c
K-1 71) =
1
w - krwTr(R, (x))
PARTIAL GlBBS DERIVATIVES
187
Functions of this kind for the particular case of the dyadic group were apparently first investigated in [37] in the Walsh-Fourier multiplier theory. Such functions were later used for the dyadic derivatives in [4] for the same purposes as those considered here. To be consistent with these particular definitions we omitted the factor g-' the appearance of which could be expected from the convolution theorem. By using the function 1V-k we introduce an inverse operator called the Gibbs anti-derivative on finite non-Abelian groups [23]
Definition 6.2 Forafunction f E P ( G )the Cihhs anti-derivative of order k, denoted by I k , is dejined by
The Gibbs anti-derivative can be considered as a Fourier multiplier operator, thus having all properties characteristic for these operators. Therefore, there is no need for any particular consideration of these properties here. Having the concept of Gibbs anti-derivative, we can deduce a theorem which shows how to determine the values of a function f from the values of its Gibbs derivative of order k .
Theorem 6.2 Let f E P ( G )be such that S f ( 0 )= 0. Then,
ol; equivalently,
f(.)
= y-lD"P
f)(.).
Here the factor 9-l appears at the right hand side of the above equations since it was omitted in the dejinition of S W - C . Note that Theorem 6.2 can be regarded as a kind of counterpart of the so-called fundamental theorem for dyadic analysis due to Butzer and Wagner [4]. Moreover, as is noted in [2], see also [21], theorems of this kind are a kind of counterpart of the fundamental theorem of the Newton-Leibniz calculus in abstract harmonic analysis. 6.3
PARTIAL GlBBS DERIVATIVES
As we noted above, a given function f E P ( G ) , G-decomposable group, can be viewed as a function of several variables f ( z 1 , . . . ,xn), z, E G,. Therefore, partial Gibbs derivative o f f with respect to the variable 5, can be defined [26]. From Definition 6.1 and some well-known properties of unitary irreducible representations
w=o
u=o
188
GlBBS DERIVATIVES ON FINITE GROUPS
From the invariance under translation of the Haar integral
it follows
w=o
u=o
which suggests the following definition.
Definition 6.3 Partial Gibbs derivative (A, f )(x) at a point x = ( X I , . . . ,x,-1, x,, x,+l,.. . , x,) E G with respect to the i-th variable x, of a furiction f t P ( G )is deJined as the Gibbs derivative (Df,)(z,), at x,, of thefunction f , ( Y ) = f(xlr.'.,x2-lrY,xz+l,.-.,5n)). Thus,
w=o
U ,
where g, is the order of G,, K , denotes the number of nonequivalent unitary irreducible representations of G,, and rb is the dimerision of the representation RL of
G,. Actually, the partial Gibbs differentiator A, thus defined is the restriction on G, of the Gibbs differentiator on G. It follows that the partial Gibbs derivatives have properties corresponding to those of the Gibbs derivative.
Theorem6.3 Let f
E
P(G),Then,
1. A, f = 0 i f f is a constant on G,, i.e., if f has the same value for each X I E G,. Moreover, A,c = 0 for any constant c E P(G).
2. A,(clfl
+ ~ 2 f i )=clAzfl
+~Azf~r
C I I C ~
E P,
f1,fz
P(G).
3. Ifthe Fourier transform o f f t P ( G ) is S J ,then that of A, f is given by
s&,(w) = B,(W)Sf(W), w = 0,. . .. K
=
1,
where B,(w)is given by the vector of order y of the form
B,(w)= [O, 1,. . . , K , - 1,o. 1,.. . ,I - 1,. . . ,o, 1,.. . , K , - 11T .
189
GlBBS DIFFERENTIAL EQUATIONS
6.4
GlBBS DIFFERENTIAL EQUATIONS
Relation (6.5) introduces the Gibbs derivative of an arbitrary complex order. The Gibbs derivative of a positive integer order n can be defined recursively by Dn+lf = D ( D n f ) , n = 1 , 2 , . . .. This permits linear equations with constant coefficients in terms of Gibbs derivatives to be defined and solved. These equations can be considered as a particular case of the generalized linear equations studied in [ 161.
Definition 6.4 A linear Gibbs discrete differential equation with constant coeficients is an equation of the form n
rn
k=O
k=O
where a k , b k are real numbers, f E P ( G ) and y is the required solution. As in the case of ordinary differential equations we get the general solution, y, of the equation (6.6) as the sum of the solution yzzof the homogeneous equation and the partial solution yzs of the inhomogeneous equation, i.e., Y = Yzz
+ Yzs.
(6.7)
In order to find yzzone looks for roots of the characteristic equation of (6.6) n
k=O
Now, we have the following theorem.
Theorem 6.4 Ifthe roots { zz}, i and belong to the set ( 0 , . . . ,K
= 0, . . . , n of the characteristic equation are distinct -
1}, then the homogeneous solution of (6.6) is
n
c
z=O j , k = l
where the constants c ; ~ depend on the boundary conditions. There are some important differences encountered in solving linear Gibbs discrete differential equations with constant coefficients compared to solving of ordinary differential equations. A homogeneous equation of order 'n does not always have n linearly independent solutions. The following statements are, in a way, often taken for granted, however, we could not find a proof for these statements, anywhere. If t roots of the characteristic equation are repetitions of the other roots, then the number of linearly independent solutions of a linear Gibbs discrete differential equation of order k is rz, ,provided that each root of the characteristic equation is in the set ( 0 , . . . , K - l}.
790
GlBBS DERIVATIVES ON FINITE GROUPS
If s of the roots are not in this set, then the number of linearly independent solutions of the given equation is Ctz:-t rf,. This is not any peculiarity of the case considered here. A corresponding statement for so-called logical differential equations with Gibbs derivatives on the dyadic group is given in [9]. Moreover, it seems that an analogous statement is valid more generally, as it is noted without proof in [ 161. To get the particular solution of (6.6) we apply the Fourier transform on both sides of (6.6), and with property 3 of Theorem 6.1 we obtain: n
m
k=O
k=O
From there, providing that equation (6.8) is compatible, that is, S f (0) = 0 for all w E { z z } , i = 0 , . . . , R , we have
P
SS,(W)= -sf ( w ) ,
Q
where m
rn
k=O
k=O
By introducing the notation
H ( w ) = r,g-l-
P
Q'
we have SY(W)= .,lH(W)Sf
(w).
(6.10)
From (6.10) by using the convolution property, the inverse Fourier transform produces the particular solution
c
g-1 Yzs(X)
=
h(u)f(zu-').
(6.I 1)
u=o
It follows that (6.6) has a general solution of the form n
6.5
m
MATRIX INTERPRETATION OF GlBBS DERIVATIVES
In this section we will consider a matrix representation of Gibbs differential operators suitable for their numerical evaluation. One of the main properties characterizing Gibbs derivatives is given by
S D f ( w ) = W S f ( w ) ,W E { O , l , ...'K - l } ,
(6.12)
MATRIX INTERPRETATION OF GlBBS DERIVATIVES
191
where S f ( w ) are the Fourier coefficients of a function f E P ( G ) ,while S D ~ ( W ) denotes the Fourier coefficients of its Gibbs derivative D,f. Moreover, a wish to have a differential operator satisfying this property motivated the introduction of the class of differential operators considered here. In this setting, the relation (6.12) is used by some authors as a starting point for defining certain Gibbs derivatives on groups in terms of formal Fourier series (see, for example, [181,[241, 1251, [401, [MI, [341). Using this approach, the Gibbs derivative on a finite group can be defined in matrix notation as follows.
Definition 6.5 The Gibbs derivative D , on a finite, not necessarily Abelian, group G of order g is dejned [15], [29] as
D, = g-'[R] o G 0 [RI-', where [R] is the matrix of unitary irreducible representations of G over P, i.e., [R]= [ai3]with aij = Rj(i), j E {0,1,. . . , g - 1}] j E {0,1,.. . , K - l},G is a diagonal ( K x K ) matrixgiven by G = d i a g ( 0 , 1 , . . . ,K - l),and [R]-' = [bsq] E {0,1,. with b,, = ~ , R L ~ (s Q ) ~ . . , K - l}, q E {0,1,. . . , g - 1). For a function f ( z ) = f ( z 1 , .. . , 5), E P we define the partial Gibbs derivative with respect to the variable xi E G as a restriction on Gi of the previously introduced Gibbs derivative on G.
Definition 6.6 Let G be representable in the form (2.4). The partial Gibbs derivative Ai with respect to the variable 3:i is dejined as: n
a, = @ A j , 2=1
with
where I(g7xg,) is a (g3 x g J ) identity matrix, and 8 denotes the Kronecker product. Note that, using the representation (2.13) for the first K non-negative integers {0,1, . . . ,K - l}, the matrix G can be cxpressed as:
where, n
Qi
=
@ 23, i=l
792
GlSBS DERIVATIVES ON FINITE GROUPS
with
where G, is a diagonal ( K , x K,) matrix given by G, = drag(0, b,, b,, . . . , b,) with b, defined by (2.13). Recall that the matrix [R]is the matrix of unitary irreducible representations of G over P. Since G is representable in the form (2.4), the matrix [R]can be generated as the Kronecker product of ( K ,x 9,) matrices [R,]of unitary irreducible representations of subgroups G,, 2 = 1,.. . n, i.e.,
.
7L
PI = @IR.LI. z=1
Thanks to the well-known properties of the Kronecker product, the same applies to the matrix [RI-', i.e, for this matrix holds n
[R]-' = @[R,]-l. ,=l
By using the representations introduced above for the matrices [R]. [R]-l and G, the matrix D, of the Gibbs derivative can be rewritten as
After a short calculation, by using well-known properties of the Kronecker product we prove the following.
Proposition 6.1 The matrix D, representing the Gibbs derivative on ajinite group
G of order y can be expressed in terms of partial Gibbs derivatives as n 2=1
where the coeficients b, are dejned by (2.8).
6.6
FAST ALGORITHMS FOR CALCULATION OF GIBBS DERIVATIVES ON FINITE GROUPS
In this section we will disclose fast algorithms for computation of Gibbs derivatives on finite groups. As is noted in 1291, the application of Definition 6.5 leads to an algorithm for the computation of Gibbs derivatives of which the complexity is obviously approximately equal to the complexity of calculation of one direct and one inverse
FASTALGORITHMS FOR CALCULATION OFGIBBS DERIVATIVES ON FINITEGROUPS
193
Fourier transform. The advantage is that the application of fast Fourier transform on groups is immediately possible without any considerable modification. However, from the computational point of view a more efficient algorithm for the computation of Gibbs derivatives on finite groups can be disclosed defining the Gibbs derivative in terms of partial Gibbs derivatives, that is, starting from the relation (6.13). Moreover, the algorithm thus obtained is quite suitable for a parallel implementation. The idea comes from the following facts. As is noted in Section 3.1, the definition of the fast Fourier transform (FFT) on group G is based upon the factorization of G into the equivalence classes relative to some subgroups of G. On the other hand, the i-th partial Gibbs derivative on a group G representable in the form (2.4) is defined as the restriction of Gibbs differentiation on G to the differentiation on G,. Therefore, it is natural to search for a fast algorithm for the computation of Gibbs derivatives through the partial Gibbs derivatives. Note that the i-th partial Gibbs derivative is defined (Definition 6.6) by a relation of the form (3.1). Therefore, comparing the matrices Ai, i E (1,. . . ,n } , for a given group G with the matrices [ C Y k ] appearing in the factorization of the Fourier transformation matrix, we infer a strong similarity. It follows that the algorithm for the computation of i-th partial Gibbs derivative will be similar to the i-th step of the FFT, and, hence, may be described by a flow-graph similar to that describing the i-th step of the FFT. Naturally, the similarity is less strong in the case of non-Abelian groups than in the case of Abelian groups, for the dual object r of a non-Abelian group G does not have the structure of a group isomorphic with G as is the case for Abelian groups. More precisely, the cardinality of r is not equal to the order of G, and the consequence is that in the case of non-Abelian groups the number of input nodes in the flow-graph is different for each step of the FFT, as we noted above, while all partial Gibbs derivatives are by definition applicable to the vector f whose order is y. Nevertheless, the similarity in the overall structure of the corresponding flow-graphs is retained and can be efficiently used for the disclosure of the fast algorithms for the computation of partial Gibbs derivatives. As we noted above, the flow-graph for the computation of the k-th partial Gibbs derivative of a function f on a finite group G of order g consists of g input and g output nodes of which some are connected by branches. It is determined by the overall structure of the matrix Ak which nodes will be mutually connected, as in the case of the FFT. More precisely, the output node j will be connected with the input node i iff the element di, of A, describing the k-th partial Gibbs derivative is not equal to zero. As in the case of the FFT, a weighting coefficient is associated to each branch. However, all weights in the fast algorithm for the computation of Gibbs derivatives are numbers belonging to P even for non-Abelian groups, which is another considerable difference relative to FFT for this case. Denoting by k ( i ,j ) the branch connecting the output node i with the input node j,the weight w'((i,j) associated with this branch is given by w ( i , j ) = d,,, where d t j is the (i,j)-th element of A,. Having the fast algorithms for the computation of partial Gibbs derivatives, one obtains the fast algorithm for the computation of the Gibbs derivative of a function f E P ( G )according to (6.13) simply by adding the output nodes of the flow-graphs
194
GIBBS DERIVATIVES ON FINITE GROUPS
for the calculation of partial Gibbs derivatives multiplied by the weight coefficients b, defined by (2.13). Now, we give a brief analysis of the complexity of the algorithm described above. The number of calculations is usually employed as a first approximation to the complexity of an algorithm. It is obvious from Definition 6.6 that the number of calculations required to calculate A, f is equal to gg,, since there are at most g, non-zero elements in each row of the matrix A,. According to (6.13), the number of operations needed to calculate the Gibbs derivative D, f is equal to g(Cr=lg,) followed by ( n - l ) g multiplications with the weighting factors b, and n g additions. Recall that the number of calculations in an FFT to which our algorithm can be compared is g(c:=l g 2 ) . However, to confirm the efficiency of the algorithm proposed it is important to give at least a rough estimate of its overall time complexity. It is obvious from Definition 6.6 that, unlike the FFT, which is a sequential algorithm in its essence for the input to one stage is the output from the preceding stage, the fast algorithm for the calculation of Gibbs derivatives (FGD) is quite suitable for a parallel implementation, since the calculation of the partial Gibbs derivatives can be carried out simultaneously. This parallelism is over and above the parallelism in each step of the calculation of each A, as in the corresponding step of the FFT. It follows that the time complexity of the FGD does not depend on the number of subgroups of G as is the case with the FFT. The FGD is considerably faster than the corresponding FFT, since it can always be implemented in only two steps, compared with the 7~ steps required in the FFT. Of course, the price is the number of processors operating in parallel and the memory storage requirements, which in this case should certainly be greater for the FGD than for the FIT. A more accurate analysis of the complexity of the FGD is certainly needed, but to be correct it may only be done after rather precise specification ofthe facilities used for the implementation. As is usually the case in the study of problems like that considered here, the procedure for the numerical calculation is best explained by some examples. We shall therefore consider two examples, the first for Abelian groups and the second for non-Abelian groups.
Example 6.1 Let G = 2, = ({0,1,2,3,4,5,6,7,8}, 0 ) be the group of nonnegative integers less than 9 with componentwise addition modulo 3 of 3-adic expansions of group elements as the group operation. For convenience the group operation is shown in Table 6.1. The group representations of Z g over the complex$eld are the Vilenkin-Chrestenson functions shown in a matrix form in Table 6.2. Note that the group 2 9 can be considered as the product 29 = 2 3 x 2 3 where 2 3 = ({0,1, a}, 3 ) is the group of non-negat66ive integers less than 3 with addition modulo 3 as group operation. Therefore, any complex-valued function f on Z g ran be considered as a two-variable function f (21,x2), x1,22 E 2 3 . The matrices A: and A; of the partial Gibbs derivatives relative to the variables x1 and x2,respectively, and the matrix D g of the Gibbs derivative on Z g calculated according to (6.13)as Dg = 3A: A: are shown in Figure 6.1 a, b , c, respectively.
+
FASTALGORITHMSFORCALCULATION OFGIBBS DERIVATIVESON FINITEGROUPS
195
Table 6.1 Group operation of Z9.
010
1
2
3
4
5
6
7
8
1 2 3 4 5 6 7 8
2 0 4 5 3 7 8 6
0 1 5 3 4 8 6 7
4 5 6 7 8 0 1 2
5 6 7 8 6 1 2 0
3 7 8 6 7 2 0 1
7 8 0 1 2 3 4 5
8 6 1 2 0 4 5 3
6 7 2 0 1 5 3 4
7 8
I
Table 6.2 The group representations of Z9 over C.
Thepow-graph of the fast algorithm for the computation of the Gibbs derivative D g of a complex-valued function f on Z g given by its truth vector f = [ f ( O ) , . . . , f ( 8 ) l T is shown in Figure 6.2.
Now, let us consider as the second example the calculation of the Gibbs derivatives of functions defined on the non-Abelian group of binary matrices described in [ 151.
Example 6.2 Let G be the multiplication group of the twelve ( 3 x 3) matrices t = [t,,], i , j = 0,1,2, over the complexfield represented in Table 6.4. For convenience the group operation is explicitly shown in Table 6.3. Note that G is isomorphic to the direct product of the cyclic group C2 = ( ( 0 ,l},0 ) of order 2 with generating
796
GlBBS DERIVATIVES ON FINITE GROUPS
-1 0 0
O O l O O b O O
a O O l O O b O
O a O O l O O b
l a b O b l a O U b l O 0 0 0 1 A:= O O O O O O a O O O O O O O O -O O O O
O O O a ~ b O O O
O O O b ~ l O O O
A:=
-4
D g =
O O b O O a O -O
l O O b O O a O
b
O l O O b O O a
a 0 0
a.
b.
b O O a O O l O O
0
b O O a O O l O
-
O b O O a O O l
,
O O O O O O O O O 0
O l b a
U
0
O a l b
3 a O 0 3 b b 4 a 0 3 a O 0 a b 4 0 0 3aO 3 b 0 0 4 a b 3 a 0 3 6 0 b 4 a 0 0 3 b a b 4 0 0 3a 0 0 3b 0 0 4 0 3 a O 0 3 6 0 b -0 0 3 a O 0 3 b a a
0
0
O O b a l
O
,
-
0 0 3 b 0 0 3b O 0 3 a O 0 3a a b 4 a
b
O
-
,
4 -
Fig. 6.7 a. The partial Gibbs derivative A: on Z s , 6. The partial Gibbs derivative A; on Z s , c. The Gibbs derivative Ds on Z9.
FASTALGORlTHMSFORCALCULATlONOFGIBBS DERIVATIVESON FINITEGROUPS
197
Fig. 6.2 The flow-graph of the fast algorithm for calculation of the Gibbs derivative Dg
On
Zg.
198
GIBBS DERIVATIVES ON FINITE GROUPS
o
0
1
2
3
4
5
6
7
8
9 1 0 1 1
0
0
1
2
3
4
5
6
7
8
9 1 0 1 1
1
1
2
0
5
3
4
7
8
6 1 1
2
2
0
1
4
5
3
8
6
7 1 0 1 1
5
3
3
4
4
5
5
3
4
1
6
6
7
8
9 1 0 1 1
7
7
8
6 1 1
8
8
6
7 1 0 1 1
9
9
9 1 0 1 1
6
7
8
9
8
6
9 1 0
7
8
4
5
1 0 1 0 1 1 1 1 1 1
3
0 2
1
0
1
2
2
9 1 0 1 1
1
0
1
1
9 1 0
6 9
7 8
6
9 8 7
9 1 0
7
8
6
1
2
3
4
5
1
2
0
5
3
4
2
0
1
4
5
3
3
4
5
0
1
2
7
4
5
3
2
0
I
6
5
3
4
I
2
0
0 1 1
9 1 0
0
element 1 arid the symmetric group of permutations S3 (see Table 2.1 and Table 2.5). Table 6.4 lists also all absolutely irreducible representations for the given group G = C2 x S3 over the Galois field GF(1I ) (GF( 1I ) is a splitting field for G). A given function f G + GF(11) can be considered as a two-variable function f ( 5 1 , 5 2 ) r x1 E {(All, 5 2 E {0,1,2,3,4,5). The matrices Ai2 and Ai2 ofpartial Gibbs derivatives with respect to the variables X I and 2 2 are shown in Figure 6.3 and Figure 6.4, respectively. For the given group G the number of unitary irreducible representations is K = K1KZ with K1 = 2 , K2 = 3, and, therefore, the matrix D12 of the Gibbs derivative may be calculated according to (6.13) as D12 = 3Ai2 + Ah2. This matrix is shown in Figure 6.5. Thepow-graph of the fast algorithm for the calculation of the Cibbs derivative of a function f mapping G into G F ( l 1 ) is shown in Figure 6.6.
6.6.1
Complexity of Calculation of Gibbs Derivatives
There are different approaches to efficiently calculate the Gibbs derivatives on finite groups. From Definition 6, the Gibbs derivatives can be rcgarded as convolution operators and, therefore, can be calculated through convolution algorithms [ 191. These algorithms can be derived in analogy to the convolution algorithms defined in terms of FFT [ 13, [ 191. For Gibbs derivatives in P(G),time and space complexity approximate to O(2n 1),and O(g), respectively, if in-place computation [35] is assumed. FFT-like algorithms for calculation of Gibbs derivatives are derived through the application of the Good-Thomas factorization [ 131, [36], to the matrix representing the Gibbs derivative on a given group G [29]. Unlike the algorithms for calculation
+
FASTALGORITHMS FOR CALCULATIONOFGIBBSDERIVATIVESONFINITE GROUPS
Table 6.4 The representations of G12 over GF(l1).
0
1
1
1
1
1
1
1
1
1
2
1
1
I
1
3
1
10
1
10
4
1
10
1
10
5
1
10
I
10
6
1
I
10
10
7
1
I
10
10
8
1
1
10
10
9
1
10
10
1
10
1
10
10
1
11
1
10
10
1
199
200
GIBBS DERIVATIVES ON FINITE GROUPS
6 0 0 0 0 0 5 0 0 0 0 0
0 6 0 0 0 0 0 5 0 0 0 0
0 0 6 0 0 0 0 0 5 0 0 0
0 0 0 6 0 0 0 0 0 5 0 0
0 0 0 0 6 0 0 0 0 0 5 0
0 0 0 0 0 6 0 0 0 0 0 5
5 0 0 0 0 0 6 0 0 0 0 0
0 5 0 0 0 0 0 6 0 0 0 0
0 0 5 0 0 0 0 0 6 0 0 0
0 0 0 5 0 0 0 0 0 6 0 0
0 0 0 0 5 0 0 0 0 0 6 0
0 0 0 0 0 5 0 0 0 0 0 6
Fig. 6.3 The partial Gibbs derivative Ai2 on G I ~ .
7 5 5 9 9 9 0 0 0 0 0 0
5 7 5 9 9 9 0 0 0 0 0 0
5 5 7 9 9 9 0 0 0 0 0 0
9 9 9 7 5 5 0 0 0 0 0 0
9 9 9 5 7 5 0 0 0 0 0 0
9 9 9 5 5 7 0 0 0 0 0 0
0 0 0 0 0 0 7 5 5 9 9 9
0 0 0 0 0 0 5 7 5 9 9 9
0 0 0 0 0 0 5 5 7 9 9 9
0 0 0 0 0 0 9 9 9 7 5 5
0 0 0 0 0 0 9 9 9 5 7 5
0 0 0 0 0 0 9 9 9 5 5 7
Fig. 6.4 The partial Gibbs derivative A;’ on G12,
CALCULATION OF GlBBS DERIVATIVES THROUGH DDS
3 5 5 9 9 9 4 0 0 0 0 0
5 3 5 9 9 9 0 4 0 0
5 5 3 9 9 9 0 0 4 0 0 0 0 0
9 9 9 3 5 5 0 0 0 4 0
9 9 9 5 3 5 0 0 0 0 4 0 0
9 9 9 5 5 3 0 0
0 0 0
4
4 0 0 0 0 0 3 5 5 9 9 9
0 4 0 0 0 0 5 3
5 9 9 9
0 0 4 0 0 0 5 5 3 9 9 9
0 0 0 4 0 0 9 9 9 3 5 5
0 0 0 0 4 0 9 9 9 5 3 5
201
0 0 0 0 0 4 9 9 9 5 3 3
Fig. 6.5 The Gibbs derivative DI2.
of Fourier transform on groups, the steps in FFT-like algorithms for Gibbs derivatives can be performed simultaneously. That approach reduces the time complexity at the price of the space complexity. Gibbs derivative on any finite group can be calculated in two steps. However, the space complexity approximates to O(ng + 9). 6.7
CALCULATION OF GlBBS DERIVATIVES THROUGH DDS
Both convolution and FFT-like algorithms for Gibbs derivatives are based upon the truth-vector representation of a given function f on G. Therefore, their complexity is determined by the order y of G. In practical applications that limits the use of these algorithms to functions of a relatively small number of variables. Algorithms based on MTDDs permit calculation of Gibbs derivatives of functions of a considerable number of variables. A procedure to calculate Gibbs derivatives is based on decomposition of Gibbs derivative into the linear combination of partial Gibbs derivatives in (6.13). It is derived as a generalization of the procedure for calculation of the Fourier transform on non-Abelian groups and as a modification of the procedure for calculation of Gibbs derivatives on finite Abelian groups through DDs [32]. The procedure consists of the following steps. Procedure f o r calculation of Gibbs derivatives 1. Represent f by the MTDD.
2. Determine partial Gibbs derivatives.
3. Determine the Gibbs derivative as the linear combination of partial Gibbs derivatives.
SdnOtlO 311NIJ NO S3AllVAIt130 Sf3BlD
ZOZ
CALCULATION OF GIBBS DERIVATIVES THROUGH DDS
203
The partial Gibbs derivatives are calculated through MTDD for f and represented again by MTDDs. The Gibbs derivative is determined by adding MTDDs representing the partial Gibbs derivatives. 6.7.1 Calculation of partial Gibbs derivatives The procedure for calculation of partial Gibbs derivatives is similar to that for FFT. The difference is in the processing rules applied at the nodes and cross points in the MTDD for f . For the partial Gibbs derivative with respect to zi,the nodes and cross points at the i-th level are processed by the rule determined by Di. The nodes and cross points at the other levels are processed by the rules determined by the identity matrices of the corresponding orders as determined in Definition 6.6. We assume that the 21 is assigned to the root node, and the other variables are assigned to the other levels in the increasing order.
Procedure for calculation of partial Gibbs derivative Di Given a function f on the decomposable group G of the form (2.4). 1. Represent f by the MTDD.
2. Process the nodes and cross points in the MTDD in a recursive way level by level starting from the nodes at the level to which z, is assigned up to the root node. 3. For j = n to 1, process the nodes and the cross points at the j-th level by using the rule determined by DG, if j = i, I(g,xg,) if j < i, and I ( K , ~ K ,if) j > i . The output from the processing of the root node is the partial Gibbs derivative o f f with respect to the variable 5,. The procedure for calculation of the Gibbs derivatives through MTDDs is explained and illustrated by the following example.
Example 6.3 The Gibbs derivative for functions on the group G24 in Example 3.5 is dejned by Df = 6D1 3 D 2 D3, where the partial Gibbs derivatives are given by D1 = D c ~8 I ( 2 x 2 ) 63 I ( 6 x 6 ) J DZ = I ( 2 x 2 ) 63 Dc, @ I ( 6 x 6 ) . D3 = I ( 2 x 2 ) @3I ( 2 x 2 ) c 3 Dsy with
+
+
-
6
5 61.
Ds3=
7 5 5 9 9 -9
5 7 5 9 9 9
5 5 7 9 9 9
9 9 9 7 5 5
9 9 9 5 7 5
-
9 9 9 5 5 7 -
’
204
GlBBS DERIVATIVES ON FINITE GROUPS
To calculate the partial Gibbs derivative with respect to 2 3 we perform calculations determined by dejinition of Ds3at the nodes q3,o. q3.1, q3,3 and the crosspoint 9 3 , ~ In . the nodes at the levels corresponding to 5 2 and 21, we perform the identical mapping dejined by I ( 2 x 2 ) . To calculate the partiul Gibbs derivative with respect to 2 2 . we perform the identical mapping determined by I ( 6 x 6 ) at the nodes and cross point at the level f o r 2 3 , the calculations determined by Dc, at the nodes f o r 2 2 and the identical mapping determined by I ( z x 2 ) at the root node. Similar; to calculate partial Gibbs derivative with respect to 2 1 , we perform Dc, at the root node, while at the other nodes and the cross points the identical mappings I ( 2 x 2 ) are performed. 1. Partial Gibbs derivative with respect to 2 3 : q3,o
=
DsJ [0,6,2,1,0,
=
[5,6,9,2,0,0IT ,
q3.1
=
Dsl [ 2 , 1 , 1 , 0 , 0 ,OIT
=
[ 2 , 0 , 0 , 3 , 3 ,31T
q3,2
=
DS,
=
[o,0,0, 0, 0 , 0 l T ,
43,3
=
D s J [I,1 , 1 , 1 , 2 ,2IT = [7,7,7,10,1,1IT.
Q~,I =
=
I(zX2)
1, 1, 1, 1,
[ ] q3,3
=
[0,0,0,0.0,0,7,7,7,10. 3,3IT
[ 5 , 6 , 9 , 2 , 0 , 0 , 2 , 0 , 0 , 3 , 3 , 3 , 0 , 0 , 0 , 0 , 0 , 0 , 7 , 7 , 7 , 11,1IT. 0,
2. Partial Gihhs derivative with respect to 2 2 :
=
[O, 0, 0, 0,5,5,0,0,0,0,6, 6]T.
CALCULATION OF GIBES DERIVATIVES THROUGH DDS
=
205
[lO,S,G,S,O,O, 1,3,5,5,0,0,0,0,0,0,5,5,0,0,0,0,6,6]T.
3. Partial Gibbs derivative with respect to 2 1 : 43,0, q3,1, q3,2,
q2,1
=
and 43,3 are as in calculation of D2,
=
10x2)
[ ] q3,2 43.3
= [ l , l l, , l , l ,1 , 1 , 1 , 1 , 1 , 2 , 2 ] ?
[5,8,G,0,5,5,S,0,O,5,l0,lO,6,3,5,0,6,6,5,0,0,6,l,l]T.
Theref0 re, D24
GD1+3Dz+D3 = [10,1,8,9,S,S,S,9,4,4,8,8,3,7,8,0,7,7,4,7,7,2,3,3]*. =
Each step of calculation can be represented through MTDDs. For example, Figure 6.7 shows calculation of the partial Gibbs derivative with respect to x3 for f .
206
GlBBS DERIVATIVES ON FINITE GROUPS
,
~
for D,
for D,
0+1+2+3+4
fig. 6.7 Calculation of the partial Gibbs derivative with respect to 1c3 for f in Example 3.5.
CALCULATION OF GlBBS DERIVATIVES THROUGH DDS
207
REFERENCES I . Burrus, C.S., Eschenbacher, R.W., “An in-place, in-order prime factor FFT algorithm”, IEEE Trans. Acoust., Speech, Signal Processing, Vol. ASSP-29, 1981, 806-816. 2. Butzer, P.L., Engels, W., Wipperfiirth, U., “An extension of the dyadic calculus with fractional order derivatives: general theory,” Comp. and Math., with Appls., 12B, NO. 5/6, 1986, 1073-1090.
3. Butzer, P.L., StankoviC, R.S., Eds., Theory andApplications of Gibbs Derivatives, MatematiEki institut, Beograd, 1990. 4. Butzer, P.L., Wagner, H.J., “Walsh-Fourier series and the concept of aderivative”, Applicable Analysis, Vol. 3,29-46, 1973.
5. Clarke, E.M., Zhao, X., Fujita, M., Matsunaga, Y., McGeer, R., “Fast Walsh transform computation with Binary Decision Diagram”, in Kebschull, U., Schubert, E., Rosentiel, W., Eds., Proc. IFIP WG 10.5 Workshop on Applications of the Reed-Muller Expansion in Circuit Design, 16.-17.9.1993, Hamburg, Germany, 82-85. 6. Cohn-Sfetcu, S., Gibbs, J.E., “Harmonic differential calculus and filtering in Galois fields, Proc. 1976 IEEE Con$ on Acoustics, Speech and Signal Processing, Philadelphia, 148-153, 1976.
7. Gibbs, J.E., “Walsh spectrometry, a form of spectral analysis well suited to binary digital computation”, NPL DES Rept., Teddington, Middlesex, United Kingdom, July 1967. 8. Gibbs, J.E., “Walsh functions as solutions of a logical differential equation”, National Physical Lab., Teddington, England, DES Report, No. I , 1969. 9. Gibbs, J.E. Millard, M.J., “Some methods of solution of linear ordinary logical differential equations”, NPL DES Rept., No. 2, Dec. 1969. 10. Gibbs, J.E., Simpson, J., “Differentiation on finite Abelian groups”, National Physical Lab., Teddington, England, DES Rept, No. 14, 1974. 1 1 . Gibbs, J.E., StankoviC, R.S., “Matrix interpretation of Gibbs derivatives on finite groups”, private communication 1989.
12. Gibbs, J.E., StankoviC, R.S., “Why IWGD - 89? a look at the bibliography on Gibbs derivatives”, in Butzer, P.L., StankoviC, R.S., Eds., Theory and Applications of Gibbs Derivatives, MatematiEki institut, Beograd, 1990, xi-xxiv. 13. Good, I.J., “The interaction algorithm and practical Fourier analysis”, J. Roy. Statist. Soc., ser. B, Vol. 20, 1958, 361-372, Addendum, Vol. 22, 1960, 372375.
208
GlBBS DERIVATIVES ON FINITE GROUPS
14. Karpovsky, M.G., “Fast Fourier transforms on finite non-Abelian groups”, IEEE Trans. Comput., Vol. C-26, No. 10, 1028-1030, 1977. 15. Karpovsky, M.G., Trachtenberg, E.A., “Fourier transform over finite groups for
error detection and error correction in computation channels”, Inform. Control, Vol. 40, No. 3, 335-358, 1979.
16. KeEkii, J.D., “On some classes of linearequations”, Publ. Inst. Math. (Beograd), 24, (38), 1978, 89-97. 17. Le Dinh, C.T., Le, P., Goulet, R., “Sampling expansions in discrete and finite Walsh-Fourier analysis”, Pror. 1972 Symp. Applic. Walsh Functions, Washington, DC, 1972,265-27 I .
18. Moraga, C., “Introduction to linear p-adic invariant systems, i n Cybernetics and Systenzs Research, Vol. 2, 121-124, Ed., R. Trappl, Vienna: Electronic Science Publ., 1984. 19. Nussbaumer, H.J., Fast Fourier Transform atid Convolution Algorithms, SpringerVerlag, Berlin, Heidelberg, New York, 1981. 20. Onneweer, C.W., “Fractional differentiation on the group of integers of a p-adic or p-series tield, Anal. Math.. Vol. 3, I 19-130, 1977.
21. Schipp, F., Wade, W.R., “A fundamental theorem of dyadic calculus for the unit square”, Applicable Analysis, Vol. 34, 1989, 203-218.
22. StankoviC, R.S., “A note on differential operators on finite non-Abelian groups, Applicable Analysis, Vol. 2 I , 3 1-4 I , 1986. 23. Stankovid, R.S., Differential Operators oti Groups, Ph.D.Thesis, Faculty of Elec. Engng., Univ. Belgrade. iii+ I27pp. 1986 (in Serbian, summary in English 8pp). 24. StankoviC, R.S., “A note on differential operators on finite Abelian groups”. Cybernetirs and Systems, 18, 22 1-23 I , 1987.
25. StankoviC, R.S., “A note on spectral theory on finite non-Abelian groups”, in Theory arid Applications ojSpectral Techniques, (Ed.), C. Moraga, Forschungsbericht 268, ISSN 0933-6192, Dortmund University, 1988. 26. StankoviC, R.S., “Gibhs derivatives on finite non-Abelian groups”, in Butzer, P.L., StankoviC, R.S., (Eds.), Theory and Applications of Cihbs Derivurives, MateniatiEki institut, Beograd, 1990, 269-297. 27. StankoviC, R.S., “Matrix interprctation of fast Fourier transform on finite nonAbelian groups”, Rex. Rept. in Appl. Math., YU ISSN 0353-6491, Ser. Fourier Analysis, Rept. No.3, April 1990, 1-31, ISBN 86-8161 1-03-8. 28. StankoviC, R.S., “Matrix interpretation of the fast Fourier transforms on finite non-Ahelian groups”, Proc. lnt. Con$ 017 Sigrid Processing. Beijitzg / 90, 22.-26.10.1990, Beijing, China, I 187- 1 190.
CALCULATION OF GIBBS DERIVATIVES THROUGH DDS
209
29. StankoviC, R.S., “Fast algorithms for calculation of Gibbs derivatives on finite groups,” Approx. Theory and Its Applications, 7 , 2, June 199I , 1 - 19. 30. StankoviC, R.S., “Gibbs derivatives”, Numerical Functional Analysis and Optimization, 15, 1-2, 1994, 169-181. 3 I . StankoviC, R.S., Stankovid, M.S., “Sampling expansions for complex-valued functions on finite Abelian groups”, Automatika, 25, 3-4, 147-150, 1984. 32. StankoviC, R.S., StankoviC, M., “Calculation of Gibbs derivatives through decision diagrams”, Approximation Theory and Its Applications, China, 1998. 33. StankoviC, R.S., StojiC, M.R., “A note on the discrete Haar derivative”, Colloquia Mathematica Societatis Janos Bolyai 49. Alfred Haar Memorial Conference, Budapest, Hungary, 897-907, 1985. 34. StankoviC, R.S., StojiC, M.R., “A note on the generalized discrete Haar derivative”, Automatika, 28, 3-4, 117-122, 1987. 35. StojiC, M.R., StankoviC, M.S., StankoviC, R.S., Discrete Transforms in Application, Second Edition, Nauka, Beograd, 1993 in Serbian. 36. Thomas, L.H., “Using a computer to solve problems in physics”, in Application of Digital Computers, Ginn, Boston, 1963. 37. Watari, Ch., “Multipliers for Walsh-Fourier series”, 7’6hoky Math. J . , 2, 16, 239-251, 1964. 38. Weiyi Su, “Gibbs derivatives and their applications,” Reports of the Institute of Mathematics, Nanjing University, 9 1-7, Nanjing, P.R. China, 199 I , 1-20. 39. Weiyi, Su, “Walsh analysis in the last 25 years”, Proc. The Fifth Int. Workshop on Spectral Techniques, 15.-17.3.1994, Beijing, China, I 17- 127. 40. Zelin, He, “The derivatives and integrals of fractional order in Walsh-Fourier analysis, with applications to approximation theory”, J. Approx. T h e o v , 39, 36 1-373, 1983.
This Page Intentionally Left Blank
7 Linear Systems and Gibbs Derivatives on Finite Non-Abelian Groups Linearity is a property very often used in providing mathematical models of physical phenomena. In that setting, linear shift invariant systems and in particular, linear convolution systems on groups are efficiently used in mathematical modeling of real life systems. For example, in that general ground linear time-invariant systems can be regarded as systems on the real group R. Similarly, linear discrete-time-invariant systems are an example of systems defined on the additive group of integers 2.The use of some other groups different from R and 2 offers some advantages in particular applications, see for example [ 131, [ 141. Differential operators are used in linear systems theory to describe the change of state of a system. The systems on R described by linear differential equations with constant coefficients in terms of the Newton-Leibniz derivative are probably the most familiar example. However, group theoretic models of systems and Gibbs derivatives on groups, in particular on the dyadic and groups Cpn,have attained some considerable attention [ 181, [151, [81. In what follows we will first give a short account of background to linear systems on groups and then discuss systems on finite non-Abelian groups described by discrete differential equations using Gibbs derivatives.
7.1 LINEAR SHIFT-INVARIANT SYSTEMS ON GROUPS In this section, we briefly discuss linear convolution systems with whose input and output signals are deterministic signals on groups, and point out the relationship between Gibbs differentiators and linear convolution systems. 211
212
LINEAR SYSTEMS ON FINITE NON-ABELIAN GROUPS
Fig. 7.1 Linear shift-invariant system.
Definition 7.1 A lineur invariant system S over a group G is defined as a quadruple S = ( U ,Y ,h, *) where the operation * is deJinedfor any u E U, y E Y as follows: (7.1)
i.e., of x
* is the operation of group convolution of two functions h, 11; 2 - l in G and
o
denotes the group operation.
-
is the inverse
Worded differently, the system S consists of the set U of input signals and the set Y of output signals defined respectively, as the mappings u : G X and y : G + Y , and the impulse function h defined as the mapping h : U + Y . If ( 7.1) is true for a given system S, and given u E U, y E Y , then that system computes the inputloutput pair (u, y). Figure 7.1 shows a general model of a linear shift-invariant system on groups. Note that from the system theory point of view, S is a linear input/output system whose input and output are defined over an arbitrary group G. By using different groups, various systems studied by several authors can be obtained. For example, if G is the dyadic group, the dyadic systems ware introduced by Pichler and further studied in a series of papers by this and by several other authors; see [ 121 for a bibliography up to 1989. For more recent result, we refer to [7], [8], [21]. The systems where input and output signals are modeled by functions mapping infinite cyclic group of integers into Galois fields of order 2q, q E N , GF(24) were considered by Tsypkin and Faradzhev [32]. A generalization of the concept was given in [23] where it was shown that both cyclic and dyadic convolution systems on finite groups can be regarded as special classes of permutation-invariant systems. In a more general setting, systems over locally compact Abelian groups were considered by Falb and Friedman [91. Some aspects of the theory were extended also to non-Abelian groups by Karpovsky and Trachtenberg [ 131, [ 141. Recall that systems over finite groups can be regarded as a special class of digital filters [28], [29] [3O] or a special class of discrete-time systems with variable structure [16] over a finite interval [0,g - 11, see also [31]. It may be said that in last few years the theory of linear invariant systems on groups has been well established by several authors, although a lot of further work is still needed. Regarding the applications of these systems note that they can be used as models of both information channels (for example to represent an encoder, a digital filter, or a Wicncr filter if K is the field of complex numbers and u is a stochastic signal), and computation channels if K is a finite field. For example, different criteria for the approximation of linear time-invariant systems by linear convolution systems
9
LINEAR SHIFT-INVARIANTSYSTEMS ON FINITE NON-ABELIAN GROUPS
213
System
/
A belian
Fig. 7.2 Classification of linear shift-invariantsystems on groups. on groups were discussed in [31]. Figure 7.2 shows a classification of linear shiftinvariant systems with respect to the domain groups for input and output signals.
7.2
LINEAR SHIFT-INVARIANT SYSTEMS ON FINITE NON-ABELIAN GROUPS
In the case of systems on finite non-Abelian groups the Definition 7.1 can be stated as follows.
Definition 7.2 A scalar linear system A over a j n i t e not necessarily Abelian group G is dejned as a quadruple ( P ( G )P, ( G ) ,h, *) where the input-output relation * is the convolution product on G,
i.e.,
So, an ordered pair ( f ,y) E P ( G ) x P ( G )is exactly then an input-output pair of
A i f f and y fulfill equation (7.2). The function h E P ( G ) is the impulse response of A. It is easy to show that the system A is invariant against the translation of input
function. By that we mean that if y is the output to f , then T'y is the output to T'f, for all T E G. Therefore, we denote the system A as a linear translation invariant (LTI) system.
214
LINEAR SYSTEMS ON FINITE NON-ABELIAN GROUPS
It is apparent that when G is the dyadic group, the Definition 7.2 reduces to the dyadic systems introduced in [ 171 and further studied in [ I81 and a series of papers of that and other authors. If G is the group Zprrwe obtain the systems studied in [4], [6], and [ 151. The dyadic and p-adic systems are closely related with Gibbs differentiators on the corresponding groups see, for example [ 1.51, [ 181. A corresponding relationship can be established between LTI systems and Gibbs derivatives on finite non-Abelian groups [24]. First of all, note that (6.5) shows that the Gibbs differentiator D k of order k is a LTI system having an impulse response h given by h = 6(’)), see [24]. The Gibbs discrete differential equation (6.6) can be interpreted as an input-output relation of a system A belonging to a linear combination of Gibbs derivatives on a finite non-Abelian group. From (6.7), the general output function of this system is represented as the sum of the zero-input response of the system y z aand the zero-state response yzshas the form identical to (7.2). Therefore, we infer that the scalar linear system A associated with (7.2) is a LTI system for which (6.7) represents an input-output-state relation and 11 is the impulse response of A to the unit impulse S(z). Since h is the inverse Fourier transform of H ( w ) ,the transfer function of A is given by (6.9). 7.3 GlBBS DERIVATIVES AND LINEAR SYSTEMS The relationship between linear convolution systems on locally compact Abelian and finite non-Abelian groups discussed above can be considered and summarized in a general setting as follows. In a general ground the Gibbs differentiator of order k of a function f E K ( G ) , which we denote by Dk f , is considered as the linear operator in K ( G )satisfying the relationship [27 1
where F denotes the Fourier transform operator in K ( G ) . In most cases ~ ( wk ), = wk, but in some cases a scaling factor should be applied, while in a few particular cases the function q differs and is related to the order of group G. For example, in the case of the extended Butzer-Wagner dyadic derivative [ I J q(w,k ) = ( w * ( w ) )where ~,
w*(w)= - y - 1 ) b a 2 2 , z=o
w, being the coefficients in the dyadic expansion of tu E P. In the case of Gibbs derivatives on Vilenkin groups [35],[36], [371, the function p ( w ,k ) is a function from the so-called symbol class S& [ 3 5 ]defined as p ( w , k ) = ( k ) 7 nm , 2 0, where (x)= max{l, l l ~ l l ) .
GIBBS DERIVATIVES AND LINEAR SYSTEMS
215
It should be pointed out in attempting to determine a relationship between Gibbs derivatives and linear convolution systems that
1. Thanks to the relation (7.3) and the convolution theorem in the Fourier analysis on groups, the Gibbs differentiator of order k can be considered as a convolution operator and, therefore, can be identified with a linear convolution system whose impulse response function h is given in the transform domain by ( F ( h ) ) ( w )= p(w, k ) . For example, in the case of the Gibbs derivative on finite not necessarily Abelian groups, as well as in the case of dyadic and groups Cpn, p(w, k ) = w k by definition and, therefore, h is the k-th Gibbs derivative of the &function defined as S(z) = 1 for z equals the unit element of G, and S(s) = 0 otherwise. 2. A considerable class of linear systems on groups can be described by Gibbs differential equations in a way resembling the use of classical differential equations with constant coefficients in the linear system theory on the real group R. In the other words, a linear Gibbs differential on a group is defined as a polynomial in the Gibbs differentiator with real coefficients. The linear Gibbs differential operators form a subset of the group convolution operators realized by the corresponding subset of the group convolution systems.
As we noted above such linear systems over dyadic groups were discussed i n [ I 11, [22], and for the infinite dyadic groups in [ 181. Recall that an extension of the theory to p-adic groups, the finite Vilenkin groups, was given in [6], [ 151. A generalization to finite non-Abelian groups was given in [24], see also [26] and for p-adic systems with stochastic signals in [8]. Note that the use of systems modeled by Gibbs differential equations in the processing of two-dimensional signals was suggested in [ 191, [20]. 7.3.1 Discussion As is known, the dyadic derivative is especially adapted to functions having many jumps and possessing just a few and short intervals of constancy. Even functions having adenumerable set of discontinuities like the well-known Dirichlet function can be dyadically differentiable on [0,1].In the case of finite groups, the Gibbs derivatives also provide a mean to differentiate functions on those groups. In one word, through the family of Gibbs differentiators, the advantage of the use of differential calculus extends to the theory of systems whose inputloutput signals are piecewise constant, or discrete functions. In order to point out some possible advantages of linear systems on groups modeled by the Gibbs differential equations, recall that the use of Fourier analysis in linear systems theory is based upon the convolution theorem and the relationship between the Newton-Leibniz derivative. Thanks to the first property, the Fourier transform maps the convolution into ordinary multiplication, while the second permits the translation of differential equations into the algebraic ones. As in many other areas, the application of Fourier analysis in linear systems theory is further supported by the existence
216
LINEAR SYSTEMS ON FINITE NON-ABELIAN GROUPS
of the fast Fourier transform, FFT, and related algorithms for efficient calculation of Fourier coefficients and some other parameters useful in practical applications. The Gibbs derivatives possesses the most of the useful properties of NewtonLeibniz derivative, except the product rule and, therefore, their role in the theory of linear systems on groups can be compared to that of Newton-Leibniz derivative in classical linear systems theory on R. At the same time, the Gibbs derivatives are efficiently characterized by the Fourier coefficients on groups. The matrices representing Gibbs derivatives are Kronecker product representable in the case of finite decomposable groups, (see Section 4.7), and, therefore, the fast algorithms for the calculation of the values of Gibbs derivatives on these groups can be defined, (see Section 4.8). It may be said that the Gibbs differentiation shares some of very useful properties of both Fourier analysis and differential calculus. Thanks to these properties the Gibbs derivatives could be very promising for the use in the theory of linear systems on groups. Some recent results and extensions of the theory are given in [7], [21].
GlBBS DERIVATIVES AND LINEAR SYSTEMS
21 7
REFERENCES 1. Butzer, P.L., Engels, W., Wipperfiirth, U., “An extension of the dyadic calculus with fractional order derivatives: general theory,” Comp. and Math., with Appls., 12B, NO. 5/6, 1986, 1073-1090. 2. Butzer, P.L., Wagner, H.J., “Walsh-Fourier series and the concept of a derivative”, Applicable Analysis, Vol. 1, No. 3, 29-46, 1973. 3. Butzer, P.L., Stankovid,R.S., Eds., Theory andApplications of Gibbs Derivatives, MatematiEki institut, Beograd, 1990. 4. Cohn-Sfetcy, S . , “On the theory of linear dyadic invariant systems”, Proc. Symp. Theory and Applic. Walsh and Other Non-sinus. Functions, Hatfield, England, 1975.
5. Cohn-Sfetcy, S . , Topics on generalized convolution and Fourier transform: Theory and applications in digital signal processing and system theory, Ph.D. Thesis, McMaster University, Hamilton, Ontario, Canada, 1976. 6. Cohn-Sfetcu, S . , Gibbs, J.E., “Harmonic differential calculus and filtering in Galois fields,” Proc. IEEE Int. Con$ Acoust. Speech and Signal Processing, Philadelphia, 1976 April 12-14, 148-153. 7. Endow, Y., “Walsh harmonizable processes in linear system theory”, Cybernetics and Systems, 1996,489-5 12. 8. Endow, Y., Stankovii, R.S., “Gibbs derivatives in linear system theory”, in: R. Trappl, Ed., Proc. Twelfth European Meeting on Cybernetics and Systems Research, 5.43.4.1994,Vienna, Austria. 9. Falb, P.L., Friedman, M.I., “A generalized transform theory for causal operators”, SIAM J. Control, 1970, 452-47 1. 10. Gibbs, J.E., Simpson, J., “Differentiation on finite Abelian groups”, NPL DES Rept., No. 16, July 1974. 1 1. Gibbs, J.E., Marshall, J.E., Pichler, F.R., “Electrical measurements in the light of system theory”, IEE Conference on the Use of Computers in Measurement, University of York, September 24-27, 1973, ii+Spp. 12. Gibbs, J.E., Stankovid, R.S., “Why IWGD- 89? A look at the bibliography of Gibbs derivatives,” in: Theory and Applications of Gibbs Derivatives, P.L. Butzer, R.S. StankoviC, Eds., MatematiEki institut, Beograd, 1990, xi - zxiv.
13. Karpovsky, M.G., Trachtenberg, E.A., “Some optimization problems for convolution systems over finite groups”, In$ and Control, 34, 1977,227-247.
218
LINEAR SYSTEMS ON FINITE NON-ABELIAN GROUPS
14. Karpovsky, M.G., Trachtenberg, E.A., “Statistical and computational performances of a class of generalized Wiener filters”, IEEE Trans. Inform. Theory, Vol. IT-32, NO. 2, 1986, 303-307.
15. Moraga, C., “Introduction to linear p-adic systems,” R. Trappl, Ed., Cybernetics and Systems Research, 2, North-Holland, Amsterdam, 1984.
16. Nailor, A.W., “A transform technique for multivariable, timevarying, discretetime linear systems”, Automatica, 2, 1965,211-234. 17. Pearl, J., “Optimal dyadic models of time-invariant systems”, IEEE Trans. Computers, Vol. C-24, 1975. 18. Pichler, F., “Walsh functions and linear system theory”, Proc. Applic. Walsh Functions, Washington, DC, 1970, 175-182. 19. Pichler, F., “Fast linear methods for image filtering”, in: D.G. Lainiotis, N.S. Tzanues, Eds., Applications of Information and Control Systems, D. Reidel Publishing Company, Dordrecht, 1980, 3-1 1.
20. Pichler, F., “Experiments with I-D and 2-D signals using Gibbs derivatives”, in: Theory and Applications of Gibbs Derivatives, P.L. Butzer, R.S. Stankovid, Eds., MatematiEki institut, Beograd, 1990, 181- 196. 2 1. Pichler, F., “Realizations of Priggogine’s A-transform by dyadic convolution”, in Trappl, R., Horn, W., (eds.), Austrian Society for Cybernetic Studies, ISBN 385206127X, 1992. 22. Pichler, F.R., Marshall, J.E., Gibbs, J.E., “A system-theory approach to electrical measurements”, Colloquium on the Theory and Applications of Walsh Functions, The Hatfield Polytechnic, Hatfield, Hertfordshire, June 28-29, 1973. 23. Siddiqi, M.U., Sinha, U.P., “Permutation-invariant systems and their application in the filtering of finite discrete data”, Proc. IEEE Itit. Con$ on Acoustics, Speech, and Signal Processing, ICASSP’77, May 1977, Vol. 2, 352-355. 24. Stankovid, R.S., “Linear harmonic translation invariant systems on finite nonAbelian groups”, R. Trappl, Ed., Cybernetics and Systems Research NorthHolland, Amsterdam, 1986.
25. StankoviC, R.S., “A note on differential operators on finite non-Abelian groups,” Applicable Anal., 2 1, 1986,31-41. 26. StankoviC, R.S., “A note on spectral theory on finite non-Abelian groups,” 3rd Itit. Workshop on Spectral Techniques, Univ. Dortmund, 1988, Oct.4-6, 1988. 27. StankoviC, R.S., “Gibbs derivatives”, Numerical Functional Analysis and Optimization, 15, 1-2, 1994, 169-181.
GlBBS DERIVATIVES AND LINEAR SYSTEMS
279
28. Trachtenberg, E.A., “Systems over finite groups as suboptimal filters: a comparative study”, P.A. Fuhrmann, Ed., Proc. 5th Int. Symp. Math. Theory of Systenzs and Networks, Springer-Verlag, Beer-Sheva, Israel, 1983, 856-863. 29. Trachtneberg, E.A., “Fault tolerant computing and reliable communication: a unified approach”, Information and Computation, Vol. 79, No. 3, 1988, 257279. 30. Trachtenberg, E.A., Karpovsky, M.G., “Optimal varying dyadic structure models of time invariant systems”, Proc. 1988 IEEE lnt. Symp. Circuits and Systems, Espoo, Finland, June 1988. 31. Trachtenberg, E.A., “SVD of Frobenius matrices for approximate and multiobjective signal processing tasks”, in: E.F. Deprettere, Ed., SVD and Signal Processing, Elsevier North-Holland, Amsterdam/New York, 1988, 33 1-345. 32. Tsypkin, Ya.Z., Faradzhev, R.G., “Laplace-Galois transformation in the theory of sequential machines”, Dokl. Akad. Nuuk USSR, Vol. 166, No. 3, 1966, 507-573 (in Russian).
33. Wade, W.R., “Decay of Walsh series and dyadic differentiation”, Trans. Amer. Math. SOC.,Vol. 277, NO. I , 413-420, 1983. 34. Watari, Ch., “Multipliers for Walsh-Fourier series”, T6hoky Math. J., 2, 16, 239-25 1, 1964. 35. Weiyi, Su., “Pseudo-differentia1 operators in Bessov spaces over local fields”, J. Approx. Theory and Appls., Vol. 4, NO. 2, 1988, 1 19-129. 36. Weiyi Su, “Gibbs derivatives and their applications,” Reports of the Institute of Mathematics, Nanjing University, 91-7, Nanjing, P.R. China, 1991, 1-20. 37. Weiyi Su, “Pseudo-differentia1 operators and derivatives on locally compact Vilenkin groups,” Science in China (Series A), Vol. 35, No. 7, 1992, 826-836.
This Page Intentionally Left Blank
Hilbert Transform on Finite Groups In theory of real variable functions the Hilbert transform is defined in the following way.
Definition 8.1 The Hilbert transform f" of a function f E LP, 1 de$ned, see f o r example [2], /9],by
#
p
#
M
is
where the notation u.p. means that the integral is understood in the sense of Cauchy principal value. The functions belonging to L2 are the most widely exploited in practice, since they represent the finite energy signals. The following holds for these functions, see for example [2], [9]. Denote by F ( w ) the Fourier transform of a function f E L2. Then, the Fourier transform of its Hilbert transform, F"(ui), is given by
F"(w) = -sign(w)F(w) a e . ,
(8.2)
where -1,
w < 0,
(8.3) The relation (8.3) can be regarded as an alternative definition of the Hilbert transform. The formula holds for p = 1 if f" E L1, in which case it holds everywhere. 221
222
HILBERT TRANSFORM ON FINlTE GROUPS
Note that there are functions from l1 whose Hilbert transforms defined by (8.2) do not belong to l l . An example is f(x) = as is noted in [2]. Recall that
&
t1.p.
( 1 / m exp (*dx)) Jz;; --03
=
Lrn
-2i lim -
t-0
Jz;r
sinrx)
dx
From there it can be written at least formally: (8.4) where F is the Fourier transform operator, * denotes the convolution product on R, and the convolution integral is understood in the sense of Cauchy principal value. In the case of periodic functions with the period equal to 27r the convolution kernel used to define the corresponding Hilbert transform is {cot ;}. The approach of defining the Hilbert transform in transform domain as the multiplication by a sign function, that is by employing the relation (8.3) was used as the starting point for the introduction of a discrete Hilbert transform, i.e., the Hilbert transform for functions on finite Abelian groups. It is important to note that definitions appearing in [3] and [ I ] [4], [ 5 ] , [6] are based upon the differently defined sign functions and they coincide only in the case of cyclic groups. Recalling that the real line R exhibits the structure of a locally compact Abelian group, it can be concluded that the Hilbert transform for real-variable functions and the discrete Hilbert transform can be considered uniformly as the Hilbert transform on Abelian groups. However, the above discussed group-theoretic approach of introducing the Hilbert transform on Abelian groups through the product in the transform domain by a suitably defined sign function, can hardly be used further to extend the concept to finite non-Abelian groups. That fact becomes obvious if we recall that unlike Abelian groups, the domain r of the Fourier transform S f of a function f on a finite not necessarily Abelian group G may not have any algebraic structure suitable to define a multiplication in it which in turn can be mapped into a convolution in the group. Therefore, we have suggested in [ 8 ] just the opposite way, we have defined a Hilbert transform on a finite non-Abelian group as the pointwise multiplication of a given function by a suitably defined sign function in the group. As was shown in [8], an analysis of the properties of the thus defined transform justifies to consider it as a proper counterpart of the Hilbert transform on R or on finite Abelian groups. Therefore, we are encouraged to suggest this "opposite way to be actually a "proper" way to define a Hilbert-like transform for functions on both Abelian and non-Abelian groups permitting the considerations of these two cases in a uniform way. Recall that the same approach was already used for Hilbcrt transform on R in some particular engineering applications as for example signal filtering. In that way two aims are reached. First, the main properties of the "classically" defined Hilbert transform on Abclian groups are preserved by the "new" transform,
SOME RESULTS OF FOURIER ANALYSIS ON FINITE NON-ABELIAN GROUPS
223
and the concept is extended to finite non-Abelian groups. Further, as it will be shown below, the same approach can be used to introduce a Hilbert-like transform for functions mapping a given finite non-Abelian group into a finite field admitting the existence of a Fourier transform. 8.1 SOME RESULTS OF FOURIER ANALYSIS ON FINITE
NON-ABELIAN GROUPS For the sake of completeness of presentation in this section we disclose several further results of Fourier analysis on finite non-Abelian groups which are somewhat restricted counterparts of the corresponding results on finite Abelian groups. Recall again that the Fourier transform SF is defined on r and, thus, cannot be regarded as a function on a group and, therefore, some of the properties valid on Abelian groups are non-existent when the group is no longer commutative. Recalling the bijection V from non-Abelian group G of order g onto the subset M = (0,. . . ,g - l} of integers adopted in this monograph, note that the natural ordering "<" in M induces an ordering upon G via the inverse mapping V-l. We keep the symbol "<"for the new ordering in G and define the following partition of
G:
POSv
= {Z E
Glx < z'}, S Y M v
= {Z E
GIT = x'}, N E G v { z E G J d< z},
where x' is the inverse of x in G, i.e., x o x' = e and (x')' = x. The symbols o and e represent the group operation and the identity of G. Notice that both the set POS and N E G depend on the bijection V. Among the functions from P ( G ) we will not the following special classes.
Definition 8.2 Let f V , we say:
E
P(G). Then, with respect to the ordering of G introduced by
1. f is even iff f (x)= f ( x ' ) ,Vx E G,
2. f is odd i f f f ( x ) = -f ( d ) VX , 3. f is actual
E
G,
ifs f (x)= 0, V z E G \ P O S v ,
4. f is coactual iff f (x)= 0 , Vx E POSv, 5. f is virtual iff f (x)= 0, 'dxE G \ N E G v , 6. f is covirtual iff f (x)= 0 ,Vx
E
NEGv,
7. f is axial iff f (x)= 0 ,Yx E G \ S Y h I v ,
8. f is coaxial iff f (x)= 0 ,Vx E S Y M v . Notice that if f is odd, then f (x)= 0, 'dzE S Y M v , and if f is axial, then it is also even.
224
HILBERT TRANSFORM ON FINITE GROUPS
Definition 8.3 For all x E G, the function signv is dejined as follows
signv(x) =
{
1, x E POSV, -1, IC E NEG”, 0, Ic € SYM”.
It becomes apparent that signv is a coaxial function.
Property 8.1 Let f E P(G). Then,
+
1. f p ( x )= i ( f ( x ) f(x’))dejines an even function,
2. f O ( x )= ;( f (2)- f(x’))deJines an oddfunction, 3.
f(.)
= fe(x)
+ fd.1.
The sign minus in Property 8.1 as well as in Definition 8.2 is understood in the sense of subtraction in the field P. For the correctness of the notation, the derivation of the following properties will be restricted to the complex functions on G. The real functions will be considered as a subclass of these functions in which case we will use the notation R ( G ) .Note that the corresponding properties can be derived for function in finite fields, but the difference which should be appreciated is that in that case the notion of imaginary unity and, consequently, the complex conjugate does not exist in the “classical” sense. Therefore, the concepts of Hermitean and skew-Hermitean matrix should be appropriately reformulated as it will be given in the corresponding section below. Proof. The statement is obvious from corresponding definitions.
Property 8.2 Let f E R(G). Then f is odd iff its Fourier transform Sf is skew-T Hermitean, i.e., ifffor each 0 # w # K - 1, Sf(w)= -Sf (w) = -S;(w), where ST denotes the transpose, the complex-conjugate, and S: the complex-conjugate transpose of S f .
sf
Proof. Assume first that f is odd, i s . , f(x) = -f(z’),Vx 0 , . . . , K - 1, we have ZEG
E
G. Then, for w
=
xEG
Since f is real we obtain
Sf(w)= (-r,g-l
f (x’)Rw) (x)= - S ; ( w ) . XtG
Conversely, assume that Sf is skew-Hermitean, i.e., S, = -S;(w). Now, since f is real-valued, for each II: E G we have K- 1
K-1
f(x) =
C TT(S~(,U))R,(Z))C T r ( S f ( w ) R , ( x ) ) =
w=o
w=o
225
SOME RESULTS OF FOURIER ANALYSIS ON FINITE NON-ABELIAN GROUPS
w=o
c
w=o
-
=
K-l
K-l
=
C
T T ( S T ( W ) R ~=(~ ) ) Tr(Sf(w)RZ(z))
w=o
w=o
1TT(S~(W)R,(Z’))- f ( ~ ‘ ) .
K-l
=
w=o
Property 8.3 Let f E R(G). Then, f is even iff its Fourier transform Sf is Hermitean, i.e., #for each 0 # w # K - 1, S f ( w ) = s ; ( ~ ) . Proof. First assume that f is even, i.e., f(x) = f(z’),Vz E G. Then for all w == 0 , . . ., K 1 ~
S f ( ~ i )=
f(z)RE(z)= T w g u 1
TWg-’
C f(z’)R;(z)
xEG
JEG
T Z€G
XEC
Thus, Sf is Hermitean. Conversely, assume that Sf is Hermitean, i.e., Sf(w)= S F ( w ) . Then, since f E R ( G ) ,we have
w=o
W=O
w=o
w=o
=
f(x’)
Property 8.4 Let f
E
+ f O ( r )Then. .
R(G)and f ( z ) = f p ( x )
226
HILBERT TRANSFORM ON FINITE GROUPS
for each 0 # w
#K
-
1 and with Properties 8.2 and 8.3 we obtain
s;w
= SfL(W)
~
S.fO(7U)>
from where the assertion follows directly. From there we have that Sf- is the Hermitean part of Sf and similarly, Sf<,is the skew-Hermitean part of Sf, which we denote by H ( S f )and s H ( S f ) ,respectively. This is a direct consequence of the linearity of Fourier transform and Properties 2 and 3.
Property 8.5 The Fourier transform of the sign function is given by
Ssign(w) = rWg-' X E
Proof.
c
(RL(z)- RW(z)).
POSV
c
Ssign(w) = r,g-l
sign(x)RL(x)
xEG
-
c c
TwY-l
(RL(X) - RL(X'))
XEPOS"
=
rWg-l
(RL(z)- R W ( 2 ) ) .
X€POSV
Corollary 8.1 1. Let f E R ( G ) be covirtual. A covirtual function has a trivial decomposition in an actual function ga = ge go and an axial function gsym. Then we have
+
2. Let f E R ( G )be coactual. A coactual function has a trivial decomposition in a virtual function gv = k , k , and an axial function gsym. Then we have
+
SH(Sf) = s k , , . Proof. Let f E R ( G )be covirtual. Define ga(x) =
i
of(x).
Yx E PO& otherwise.
+
Obviously, ga is actual and gsym is even. Moreover, f = ga gsym, i.e., f = go g P .qsym. Since the sum of two even functions is even, then go represents the odd part o f f and the assertion follows from property 2 . Similarly if f is a coactual function.
+ +
HILBERT TRANSFORM ON FINITE NON-ABELIAN GROUPS
8.2
227
HILBERT TRANSFORM ON FINITE NON-ABELIAN GROUPS
As we noted above, the definition of the Hilbert transform based upon the relation (8.3) cannot be extended to functions defined on a non-Abelian group because the domain r of the Fourier transform S f may not exhibit a suitable algebraic structure. For that reason we will use a reverse approach which leads to a definition holding uniformly for Abelian and non-Abelian groups.
Definition 8.4 The Hilbert transform f" of a function f E P(G),where G is not necessarily an Abelian group, is dejned under a given ordering bijection V as the linear operator P ( G ) 4 P ( G )given by N:
f"(x)
=
- i s i g n v ( z ) f ( z ) , Vx
E
G.
The main properties of the thus defined Hilbert transform are given in the following theorem which justifies to consider it as a proper counterpart of the Hilbert transform on R and on finite Abelian groups.
Theorem 8.1 The main properties of the Hilbert transform are I . For each f,h E C(G),and a , b E C (af bh))"(z)= u f " ( x ) bh"(z).
+
+
2. I f f E R(G),then f" is purely imaginary. Moreover; i f f is even, then f is odd, and iff is odd, f " is even. __
3. For each f E C ( G ) , ( f ( x ) ) - = - f - ( x ) , where f ( x ) denotes the complex conjugate of f ( x ) . 4. For each f ,h E C ( G ) , f " ( x ) h ( z )= f ( x ) h " ( z ) 5. The inverse Hilbert transform of a coaxial function f E C ( G ) is given by
f " ( f " ( x ) ) = - f ( x ) . Notice that the actual and virtual functions are also coaxial functions, and it follows f ( x ) h ( x )= - f " ( z ) h " ( z ) .
6. Let fa E R(G) be actual. Then the Hermitean and skew-Hermitean parts of its Fourier spectrum S f are related by the Hilbert transform as shown below: For each w E (0,. . . , K - l}
WSf,,(w)) = (((isH(S.f" )
)
)
W
T
1
(8.5)
228
HILBERT TRANSFORM ON NNITE GROUPS
where T and Idenote the direct and the inverse Fourier transform, respectively.
7. Let f t R(G)be covirtual. Then equation (8.7) is also valid. Let f E R(G)be coactual. Then equation (8.9)is also valid. 8. Let f E R(G).Moreover, let Ev(f )and Od(f )denote the even and odd parts o f f , respectively. These parts are related by the Hilbert trunsform as follows: I f f is actual, then
and i f f is virtual, then
Ev(f (.)
=
(8.1 I )
-i(Od( f (.)))-,
9. I f f E R(G)is covirtual, then (8.11) is also valid. I f f E R ( G )is coactual, then (8.12) is also valid.
Proof. The following parts of the proof are numbered as the assertions of the theorem. 1 . This assertion follows directly from definition of the Hilbert transform.
2. Proof of the first part of the statement follows from definition of the Hilbert transform since the function sign is real. The second part follows from Properties 5 and 3, respectively, since sign is an odd function.
3. f"
=
-isign(.)(J) = isign(.)f = -(-isign(.)f)
=
-7.
4. Proof follows directly from definition of the Hilbert transform.
5. From definition 2 follows that sign2(.) is the identity coaxial function. It becomes apparent that since sign(.) = sign2(.) = 0 for each x E SYAlv, no inverse Hilbert transform exists for functions other than coaxial. Recall, however, that both actual and virtual functions are special kinds of coaxial functions. 6. Since f a is axial, from property I it can be written as the sum of ge and go. Then from property 6 we have
HILBERT TRANSFORM ON FINITE NON-ABELIAN GROUPS
229
and similarly
The proof is analogous in the case of virtual functions.
+
7. Let f be covirtual. Then f may be expressed as f = fa gsym, where f a is actual and gsym takes the same values as f for each z E SYhlv and is Lero otherwise. It becomes apparent that for each z E G, sign(x)gsym(z) = 0, since gsym is axial (and therefore also even). Particularly, ( g s ~ m ) ~=Y0.~ It follows
Similarly for f a coactual function.
+
+
8. Since fa is actual, then fa = ge go = E u ( f a ) O d ( f a ) .From property 1 follows that g e ( z ) = sign(z)g,(z), hence:
- w f a ( z ) ) = sign(z)Od(fa(z)) =
(-isign(z))(iOd(f,(z))) = ( i O d ( f a ) z ) ) ) -
9. See the proof of assertion 7.
Consider the following example for the illustration of the assertions of the theorem.
Example 8.1 Let G be the quaternion group Qz defined in Example 2.3. Note that the function f considered in this example is an actual function and, therefore, will be denoted here by f a . In Table 8.1 we list the values of the signv function on G and illustrate the decomposition of the given f a into an even function ge and an odd function go by using assertion I of the Theorem 8.1. Their Fourier transforms Sgcand Sgc,are given in Table 8.2. It is apparent that Sg+= H ( S f U and ) Sgc> = sH(Sf,L) as is stated in assertion 6 of the Theorem 8.1.
230
HILBERT TRANSFORM ON FINITE GROUPS
Table 8.1 The even and odd parts of the test function.
x
5’
0 I 2 3
0 3 2 1
4 5 6 7
Signv(2) 0
6 7 4 5
S Y M v = {0,2},
I 0 -1
I 1
-1 -1
yo(.)
fa(2)
0
0
0
ff
a -
a -
0
0
0
0
a -
a _
P
x
0 0
2
2
2
_
2
-B
?
-
x
-
P
-P
x
_ -x2
2
2 2
-
2
2
x
2
2
POSv = { 1 , 4 , 5 } , N E G v = { 3 , 6 , 7 }
Table 8.2 Fourier spectrum of the test function over C.
HILBERT TRANSFORM IN FINITE FIELDS
231
8.3 HILBERT TRANSFORM IN FINITE FIELDS In this section we will consider the definition of the Hilbert transform on finite nonAbelian groups for functions taking their values in a finite field admitting the existence of a Fourier transform. Notice that some of the results from Section 8.2 have to be slightly modified so that they hold also in finite fields. However, some other no longer exists in this case. The main difference which should be appreciated is that the operation we denote by * simply reduces to the transposition. In this setting the concepts of Hermitean and skew-Hermitean matrix should be reformulated.
Definition 8.5 For the Fourier spectrum Sf of afunction f E P( G )we define thefield Hermiteanpart by f H ( S j ) = S f
with signv(.) as in Definifion 8.2. It can be shown that except for Property 8.3, all other properties from Theorem 8.1 hold also in this case omitting simply thc imaginary unit. Therefore, we have the following theorem.
Theorem 8.2 The main properties of the Hilbert transform for functions belonging to P ( G )are I . For each f ,h E P ( G ) ,and a , b E P ( af b h ) ) " ( x )= a f " ( x ) + bh"(x).
+
2. I f f E P(G),is even, then f" is odd, and i f f is odd, f
is even.
3. For each f , h E P ( G )f, " ( x ) h ( z )= f ( z ) h " ( z )
4. The inverse Hilbert transform of a coaxial function f E P ( G )is given by f "( f " ( z ) ) = -f (x). Notice that the actual and virtual functions are also coaxial functions, and it follow~s f (z)h(x)= -f " ( z ) h " ( s ) . 5. Let f a E P ( G ) be actual. Then the Hermitean and skew-Hermitean parts of its Fourier spectrum S j are related by the Hilbert transform as shown below: For each w E ( 0 , . . . , K - 1)
fH(Sf<&(W)) = (((fsH(Sf,,)))L)")T,
(8.13)
232
HILBERT TRANSFORM ON FINITE GROUPS
(8.14)
Let f E P ( G ) be virtual. Then, (8.15)
(8.16) 6. Let f E P ( G )be covirtual. Then equation (8.7) is also valid. Let f E P ( G ) be coactual. Then equation (8.8) is also valid.
7. Let f E P ( G ) .Moreover. let E w ( f )and Od(f ) denote the even and oddparts o f f , respectively. These parts are related by the Hilbert trarisform as follows: I f f is actual, then
arid if f is virtual, then
Od(f ( x ) )= -(Eu( f ( 2 ) ) ) " .
(8.20)
8. I f f E P ( G )is covirtual, then (8.11) is also valid. I f f E P ( G )is coactual, then (8.12) is also valid. It follows from the properties stated in this theorem that the transform introduced by Definition 8.5 can be regarded as the Hilbert transform for functions on finite non-Abelian groups into finite fields representing a proper counterpart of the Hilbert transform introduced by Definition 8.3 and, further, as a counterpart of the classical Hilbert transform in L2 as well as the Hilbert transform on finite Abelian groups, see for example, [ I I, [4]. Example 8.2 Let G be the group described in Example 6.2. In Table 8.3 the sign function, an actual function and its even and odd parts are shown. In Table 8.4 the Fourier spectrum of thisfunction and in Table 8.3 of itsfield Hermitean andfield skewHermitean parts are given. We dejne the$eld Hermitean andjeld skew-Hermitean partsofthe Fourierspectrumrespectivelyby.fH(Sfu = Sf,Ae, and f s H ( S f a= Sf,L,,, wheref,,(z) = ~ ( f ( . ) + f ( X ' ) ) a n d f , , ( z ) = f(d)). Note that (Sf,L)T = S f u c- Sfur,and that formulas (8.15) and (8.I h ) are true.
i(f(x)-
HILBERT TRANSFORM IN FINITE FIELDS
233
Table 8.3 The even and odd parts of the test function.
z
x'
0
0
1 2 3 4 5 6
signv(x) 0
2 1 3 4 5 6
7 8 9 10
10
11
11
fae
0
0
fa0
0
a
0
q cr2
0
0 0 0 0
0 0 0 0
0
$
P
0 0
P
1
10 0
0 0
2
a __
2
0 0
w
Sf
0
Q+P
1
a+@
0 2
L!2
P __ 2
0
0
0 0
0 0
0 0 0
0
Table 8.4 Fourier spectrum of the test function.
a -
1 10
0
8 7 9
fa
5rr+5p
3 4
l001+10/3
Q
+ lop + lop
[ lOcu:+p Q
Table 8.5 Fourier spectrum of the test function in GF(11).
+p
0
(1
1
a+P
3 4
5
[
1001
CI
[
01
+ log + lop
1001
+ lopO ]
+ pO ]
0 0
[
:n+w
0
!
[
10atp 6 t r f 5 P 5a+6/!3 0
1
Gcy
!
6~+5p
5 a + 6 / 3 1001+p
+ Gp
0 0
234
HILBERT TRANSFORM ON FINITE GROUPS
REFERENCES I . Bljumin, S.L., Trakhtman, A.M., “Discrete Hilbert transform on a finite intervals”, Radiotehnika in Elektronika, No. 7, 1977, 1390-1398 (in Russian). 2. Butzer, P.L., Nessel, R.J., Fourier Analysis and Approximation, Birkhauser Verlag, Basel and Stuttgart, 197 1. 3. Ciiek, V., “Discrete Hilbert transform”, IEEE Trans., Vol. AU-18, No. 4, 1970, 340-343. 4. Moraga, C., Salinas, L., “On Hilbert and Chrestenson transform on finite Abelian groups”, Proc. 16th Int. Symp. on Multiple-Valued Logic, Blacksburg, VA, 1986.
5. Moraga, C.; Salinas, L., “On Hilbert and Chrestenson transforms on finite Abelian groups”, Scientia, 1988. 6. Moraga, C., “On the Hilbert and Zhang-Hartley transform in finite Abelian groups”, Journal of China Institute of Communications, (8), 3, 1987, 10-20. 7. Moraga, C., StankoviC, R.S., “Hilbert transform i n finite fields”, Proc. XXXVff Con& for ETAN, 20.-23.9.1993, Beograd, 59-64. 8. StankoviC, R.S., Moraga, C., “Hilbert transform on finite non-Abelian groups”, in Y. Baozong, Ed., Proc. Int. Con& on Signal Processing/Beijing ’90,International Academic Publishers, Beijing, China, 1990, 1 179- 1 182. 9. StankoviC, R.S., StojiC, M.R., Bogdanovid, S.M., Fourier Representation of Signals, NauCna knjiga, Beograd, 1988, in Serbian. 10. Vlasenko, V.A., Shodin, O.I., Microprocessors Systems for Quality Control of Digital Devices, Tehnika, Kiev, 1990 (in Russian).
Index
A Abstract harmonic analysis, I 1 Algebraic structure, 3 Arithmetic-Haar transform, 147 Arithmetic expressions, 139
B
BDD reduction rules, 77
C
Decision diagram Fourier on finite Non-Abelian groups (FNADD), 94 Fourier with preprocessing, 1 18 Functional, 86 Kronecker, 86 pseudo-Kronecker, 86 Decision diagrams (DDs), 64 Decision diagrams Fourier, 78 Multi-terminal Binary (MTBDD), 64 Multiple-place (MDD), 64 Arithmetic transform (ACDD), 86 Binary Decision Diagram (BDD), 64 Multi-terminal (MTDD), 64 Walsh (WDD), 86 Decision tree (DT), 78 Decision tree Fourier on finite non-Abelian groups (FNADT), 94 Fourier, 88 Fourier with preprocessing, I18
Differential operators, 165 Discrete Fourier transform (DFT), 37
D
Expressions Arithmetic-Haar, 147 arithmetic, 141 Arithmetic on non-Abelian groups, 146 Fixed polarity arithmetic-Haar, 15 I Fixed polarity Fourier, I50 Kronecker, 148 Reed-Muller, 144 Walsh, 143 F Fast Fourier transform (FFT), 37 Fast Fourier transform, 27 Fourier analysis, 18 Fourier transform, 19 Function actual, 205 axial, 205 coactual, 205 coaxial, 205 covirtual, 205 even, 205 virtual, 205 G Gibbs anti-derivative, 168-169 Gibbs derivatives on finite non-Abelian groups, 166 Gibbs derivatives on groups, 165
235
236
INDEX
Gibbs differentiation, 166 Gibbs discrete differential equations, 17 I Good-Thomas FFT, 65 Group character, IS Group representation theory, I 1 Group Abelian, 2 dyadic group, 2 locally compact Abelian, 2 non-Abelian, 2 representation, I2 H Hilbert transform, 203
I
Irreducible representations, I3
J
K Kronecker decision diagrams (085 L Linear shift-invariant system, 193 Linear time-invariant system, 193 M Mathematical modeling, 2 Mathematical models, 1 Matrix-valued functions, 1 IS Matrix-valued MTDD (mvMTDD), 70
N
Number-valued MTDDs (nvMTDDs), 70 0 Orthonormal system, I8
P
Parseval relation, 18 Parseval theorem, 25 Partial Gibbs derivatives, 170 Peter-Weyl theorem, 18
Q
Quaternion group, 28 R Representation
equivalent, I3 irreducible, 13 left regular, I7
S Sampling theorem, 2 Shared binary DDs (SBDDs), 72 Signals, I Signals analog, 2 continuous, 2 digital, 2 discrete, 2 Symmetric group of permutations, 21 System p-adic, 196 Dyadic, 196 Linear, 196 T Transform Arithmetic-Haar, I47 Transform discrete cosine, 38 Haar, 38 Transform Hilbert, 203 Hilbert on finite fields, 2 13 Hilbert on finite non-Abelian group, 209 Transform Vilenkin-Chrestenson. 38 U Unitary matrices, I2 Unitary representation, 12 V W Walsh expressions, 143 Walsh transform, 33 A
1,
Y 2